Embed Size (px)
Citation preview
Pure RSSI Based Low-Cost Self-localization System for ZigBee WSN
by
Philip Lin
B.A.Sc., Simon Fraser University, 2011
THESIS SUBMITTED IN PARTIAL FULFILLMENT OF
THE REQUIREMENTS FOR THE DEGREE OF
MASTER OF APPLIED SCIENCE
In the
School of Engineering Science
Faculty of Applied Science
© Philip Lin 2011
Simon Fraser University
Summer 2011
All rights reserved. However, in accordance with the Copyright Act of Canada, this work may be reproduced, without authorization, under the conditions for "Fair Dealing." Therefore, limited reproduction of this work for the purposes of private study, research, criticism, review and news reporting is likely to be in
accordance with the law, particularly if cited appropriately.
ii
Approval
Name: Philip Lin
Degree: Master of Applied Science
Title of Thesis: Pure RSSI based Low-cost Self-localization System for ZigBee WSN
Examining Committee:
Chair: Dr. Parvaneh Saeedi
Assistant Professor, School of Engineering Science
______________________________________
Dr. Bozena Kaminska
Senior Supervisor
Professor, School of Engineering Science
______________________________________
Dr. Mehrdad Moallem
Supervisor
Associate Professor, School of Engineering Science
______________________________________
Dr. Kamal Gupta
Internal Examiner
Associate Professor, School of Engineering Science
Date Defended/Approved: __August 2, 2011_________________________
iii
Abstract
In modern wireless sensor networks (WSN) applications, location awareness
has been one of the features that attracted many research interests. Various
applications utilize location information for surveillance and asset tracking
purposes. Common WSN localization systems use radio frequency (RF), ultra-
sound, or laser devices to provide range information, whose node positions are
to be determined by various algorithms accordingly. Multi-dimensional scaling
(MDS) is one of the most common algorithms for transforming inter-node
distances into node positions in Cartesian coordinates. However, MDS algorithm,
by nature, has a cubic computational complexity. Also, the algorithm‟s ability to
localize is restricted to fully connected WSNs, where every node sees every
other node. This thesis proposes a low-cost pure RF based localization system,
implemented with a novel clustering MDS algorithm. Its most attractive feature is
its ability to localize a partially connected WSN with a linear computation
complexity without sacrificing the localization accuracy. In this thesis, we review
various localization techniques and conduct experiments to compare the
clustering MDS‟ performance against the classical MDS‟ and GPS‟. The
localization with a commercial GPS, although, better than the above two
methods, has also introduced significant discrepancy. At the end, we have
demonstrated that RF localization in our low-cost system does not deliver GPS-
grade accuracy, but its ability to localize partially-connected WSN and low
computation complexity have outperformed the classical MDS approach.
iv
Acknowledgement
I would like to thank the following people for the contribution to this thesis.
Throughout my master‟s study, with the help of these individuals, I have obtained
valuable knowledge and skills to advance to the next stage in my life.
I would like to thank my supervisor, Dr. Bozena Kaminska for accepting
me into such a strong and renowned research lab and for providing me with
abundant lab resources and knowledgeable research fellows, Dr. Jens Wawerla
and Dr. Marcin Marzencki. With your valuable guidance and advice, I have had
the best school years in my life.
I would also like to thank my examining committee, Dr. Mehrdad Moallem
and Dr. Kamal Gupta, for taking part in my thesis defense and for providing
valuable suggestions.
My special appreciation goes to my research colleagues, Jian Guo and
Thomas Cho, who have contributed their personal time to help me conduct field
test, gathering data for my research.
Last, my sincere appreciation will go to my all my family member and my
girl friend, Christine Tseng, who have been supporting me throughout my
studies. Without them, I will not be able to enjoy my research years as much as I
have.
v
Table of Content
Approval ................................................................................................................ ii
Abstract ................................................................................................................. iii
Acknowledgement ................................................................................................ iv
Table of Content ................................................................................................... v
List of Tables ....................................................................................................... vii
List of Figures ..................................................................................................... viii
List of Abbreviations ............................................................................................. xii
Chapter 1 - Introduction .................................................................................... 1
1.1 Ranging Problems in Localization Systems ......................................... 1
1.2 The Self-Organizing and Self-Calibrating (SOSC) WSN ...................... 4
1.3 Scope and Outline of Thesis ................................................................ 5
Chapter 2 - Background information ................................................................. 8
2.1 ZigBee Communications Overview ...................................................... 8
2.2 MDS Algorithm ................................................................................... 14
2.2.1 Dissimilarity Matrix ................................................................... 15
2.2.2 Eigen-decomposition for MDS ................................................. 16
2.3 Iterative Closest Point (ICP) Algorithm ............................................... 18
2.3.1 Rotational Matrix...................................................................... 19
2.4 Radio Frequency Ranging Techniques .............................................. 21
2.4.1 Time Difference of Arrival (TDOA) ........................................... 22
2.4.2 Receiver Signal Strength Indicator (RSSI) .............................. 24
Chapter 3 - Related Work ............................................................................... 26
3.1 Range Based Localization Systems ................................................... 26
3.1.1 RSSI Based Localization Systems .......................................... 27
3.1.2 Non-RSSI Range Based Localization Systems ....................... 30
3.1.3 Hybrid Range Based Localization System ............................... 33
3.2 Range-free Localization Systems ...................................................... 34
3.2.1 DV-Hop Localization Algorithm: ............................................... 34
3.2.2 The Approximate Point-In-triangulation Test (APIT) Approach 36
3.2.3 The PDM Localization: ............................................................ 38
3.3 Chapter Summary .............................................................................. 39
vi
Chapter 4 - SOSC System Architecture .......................................................... 41
4.1 System / Network Architecture ........................................................... 41
4.2 SOSC System Hardware Architecture ............................................... 43
4.3 SOSC Software Architecture .............................................................. 49
4.3.1 SOSC Firmware: ..................................................................... 49
4.3.2 PC Software Architecture: ....................................................... 56
4.4 Chapter Summary .............................................................................. 62
Chapter 5 - Localization and Mapping Design ................................................ 64
5.1 Range Modeling Unit (RMU) Distance Conversion ............................ 64
5.2 SOSC RMU Architecture .................................................................... 67
5.3 Performance Comparison of the RMU ............................................... 70
5.4 SOSC Mapping Technique Design .................................................... 74
5.4.1 SOSC Map Reconstruction ..................................................... 74
5.4.2 Clustering MDS Localization Algorithm ................................... 80
5.5 Chapter Summary .............................................................................. 84
Chapter 6 - Results and Discussion ................................................................ 86
6.1 Computation Complexity .................................................................... 87
6.2 Localization Performance ................................................................... 88
6.3 Experimental Localization .................................................................. 94
6.3.1 Indoor Localization .................................................................. 94
6.3.2 Outdoor Localization ................................................................ 98
6.4 Chapter Summary ............................................................................ 115
Chapter 7 - Conclusion ................................................................................. 116
7.1 Future Work ..................................................................................... 117
References ....................................................................................................... 119
vii
List of Tables
Table 2-1: ZigBee Network Addressing Components ......................................... 11
Table 2-2: Inter-city Distance Matrix ................................................................... 14
Table 4-1: Table of Required Functional Units on each Type of Device ............. 44
Table 5-1: Average % Error and Variance of Distance Conversion with 2-meter
separation ........................................................................................................... 72
Table 5-2: Average % Error and Variance of Distance Conversion with 7-meter
separation ........................................................................................................... 73
Table 5-3: Dissimilarity Matrix of a Simple SOSC WSN ...................................... 76
Table 6-1: Position Estimation Error Comparison ............................................... 90
Table 6-2: Positioning Error of Localization ........................................................ 95
Table 6-3: Positioning Error for Different Anchor Nodes ..................................... 96
Table 6-4: Tape Measured Dissimilarities of SOSC Nodes ................................ 99
Table 6-5: Average GPS Localization Error ...................................................... 102
Table 6-6: Classical MDS Localization Accuracy .............................................. 105
Table 6-7: Clustering MDS Localization Accuracy ............................................ 107
Table 6-8: Two-sample T-Test of Statistical Significance ................................. 111
Table 6-9: Unsymmetrical Dissimilarity ............................................................. 112
viii
List of Figures
Figure 2-1: ZigBee Mesh Networking .................................................................... 9
Figure 2-2: ZigBee Mesh Networking Discovering Path ...................................... 10
Figure 2-3: ZigBee Broadcasting Mode .............................................................. 12
Figure 2-4: ZigBee Uni-casting Mode ................................................................. 13
Figure 2-5: Vancouver City Layout ...................................................................... 15
Figure 2-6: Rotational and Translational MDS Solution ...................................... 19
Figure 2-7: Translated MDS Solution and its Angle separation with the Actual
Layout ................................................................................................................. 21
Figure 2-8: TDOA Triangulation based Localization ........................................... 23
Figure 2-9: Time Traveled between Mobile Node and its Neighbouring Anchors23
Figure 2-10: Basic Radio Communication System .............................................. 24
Figure 2-11: RSSI vs. Distance Characteristic Curve ......................................... 25
Figure 4-1: System and Network Architecture .................................................... 41
Figure 4-2: Hardware Architecture of a Gateway Device .................................... 43
Figure 4-3: SOSC Gateway Device .................................................................... 45
Figure 4-4: SOSC Router Device ........................................................................ 46
Figure 4-5: SOSC End Device ............................................................................ 46
Figure 4-6: SOSC Antenna Radiation Pattern [45] ............................................. 48
Figure 4-7: Firmware Architecture for SOSC Devices ......................................... 49
Figure 4-8: RSSI Request MSC between Z-Stack Layers .................................. 50
Figure 4-9: RSSI Request Broadcast Packet ...................................................... 51
Figure 4-10: Receiving Node MSC between Z-Stack Layers .............................. 51
ix
Figure 4-11: RSSI Forward Packet ..................................................................... 52
Figure 4-12: Network Message Flow Sequence ................................................. 52
Figure 4-13: Flow Chart of SOSC Router/End Devices ...................................... 53
Figure 4-14: UART Packet Format ...................................................................... 54
Figure 4-15: Flow Chart of Gateway Firmware ................................................... 55
Figure 4-16: PC Application Software Architecture ............................................. 56
Figure 4-17: Data Manager Algorithm in the PC Application ............................... 57
Figure 4-18: Data Manager Process for Constructing the Dissimilarity Matrix .... 58
Figure 4-19: MDS Algorithm for Position Computation ....................................... 59
Figure 4-20: Position Transformation Process .................................................... 60
Figure 4-21: SOSC GUI Architecture .................................................................. 61
Figure 4-22: The Actual GUI Implemented ......................................................... 62
Figure 5-1: Experimental RSSI to Distance Relationship .................................... 65
Figure 5-2: SOSC Self-Calibrating Distance Conversion Unit Architecture ......... 68
Figure 5-3: RMU Radio Propagation Model Forming Process ............................ 69
Figure 5-4: Indoor Experimental Layout for RMU ............................................... 70
Figure 5-5: Self-calibration vs. Empirical RSSI-Distance Conversion at 2m Apart
............................................................................................................................ 71
Figure 5-6: Self-calibration vs. Empirical RSSI-Distance Conversion at 7m Apart
............................................................................................................................ 73
Figure 5-7: Map Reconstruction Procedure Flow ................................................ 75
Figure 5-8: Original Layout of the SOSC Network .............................................. 76
Figure 5-9: Map Reconstruction Step One - Translation ..................................... 77
x
Figure 5-10: Reconstruction Step Two - Rotation ............................................... 78
Figure 5-11: Reconstruction Step Three – Flipping............................................. 78
Figure 5-12: Reconstruction Step Four – Finalizing Mapping ............................. 79
Figure 5-13: Clustering MDS Algorithm .............................................................. 82
Figure 5-14: RSSI Error vs. Ranging Error ......................................................... 84
Figure 6-1: Classical MDS Map Reconstruction ................................................. 88
Figure 6-2: Clustering MDS Algorithm Map Reconstruction ................................ 89
Figure 6-3: Classical vs. Clustering MDS Localization with 13% Range Error .... 92
Figure 6-4: Classical vs. Clustering MDS Localization with 23% Range Error .... 93
Figure 6-5: Indoor Localization Setup ................................................................. 94
Figure 6-6: Localization Results from Clustering MDS and Classical MDS ........ 95
Figure 6-7: Localization Results from Clustering MDS and Classical MDS using
Different Anchors ................................................................................................ 96
Figure 6-8: Node Layout on the Field .................................................................. 98
Figure 6-9: High Density Ground Truth Layout ................................................. 100
Figure 6-10: Medium Density Ground Truth Layout .......................................... 100
Figure 6-11: Low Density Ground Truth Layout ................................................ 100
Figure 6-12: High Density GPS Localization Comparison ................................. 101
Figure 6-13: Medium Density GPS Localization Comparison ........................... 101
Figure 6-14: Low Density GPS Localization Comparison ................................. 101
Figure 6-15: RSSI vs. Distance (High Density) ................................................. 103
Figure 6-16: RSSI vs. Distance (Medium Density) ............................................ 103
Figure 6-17: RSSI vs. Distance (Low Density) .................................................. 103
xi
Figure 6-18: High Density Classical MDS Result .............................................. 104
Figure 6-19: Medium Density Classical MDS Result ......................................... 104
Figure 6-20: Low Density Classical MDS Result ............................................... 104
Figure 6-21: High Density Clustering MDS Result ............................................ 106
Figure 6-22: Med Density Clustering MDS Result ............................................ 106
Figure 6-23: Low Density Clustering MDS Result ............................................. 106
Figure 6-24: Localization Accuracy vs. Node Density ....................................... 107
Figure 6-25: Localization Error vs. Range Error (Classical MDS) ..................... 109
Figure 6-26: Localization Error vs. Range Error (Clustering MDS) ................... 109
Figure 6-27: Classical MDS Localization Error Distribution ............................... 110
Figure 6-28: Clustering MDS Localization Error Distribution ............................. 110
Figure 6-29: Localization Accuracy in Disconnected Network .......................... 113
Figure 6-30: Localization Accuracy in Fully Connected Network ...................... 114
xii
List of Abbreviations
AF Application Framework
AOA Angle of Arrival
APIT Approximate Point-In-triangulation Test
APS Ad-hoc Positioning System
BS Base Station
CID Cluster ID
CMDS Classical Multi-dimensional Scaling
COG Centre of Gravity
DV Distance Vector
dwMDS Distributed Weighted MDS
EM Electromagnetic
GPS Global Positioning System
HMM Hidden Markov Model
ICP Iterative Closest Point
LOS Line of Sight
LS Least Squared
MAP Maximum a Posteriori
MDS Multi-dimensional Scaling
MLE Maximum Likelihood Estimator
MS Mobile Station
MSREs Square Range Errors
OSAL OS Abstraction Layer
PAN Personal Area Network
PDF Probability Distribution Function
PDM Proximity Distance Map
RF Radio Frequency
RMU Range Modeling Unit
RSSI Received Signal Strength Indicator
SOC System-on-a-chip
SOSC Self-organizing and Self-calibrating
SPE Stochastic Proximity Embedding
SVD Singular Value Decomposition
TDOA Time Difference of Arrival
TOA Time of Arrival
TSVD Truncated Singular Value Decomposition
WLS Weighted Least Squared
WSN Wireless Sensor Networks
1
Chapter 1 - Introduction
1.1 Ranging Problems in Localization Systems
With the advancement in wireless technologies, wireless sensor networks (WSN)
have recently been widely researched in areas of context-aware and location-
aware applications. Context-awareness and location-awareness for a WSN
provide the network with the ability to determine ever-changing conditions in a
real-time environment. One of the most discussed and researched area in the
above two types of WSN applications is localization. The sole purpose of
localization in WSN is to facilitate network installation and to perform network
monitoring at any given area of interest within the network perimeter. The
localization ability granted to sensor nodes within the network allows the flexibility
of reconfiguring node layouts to accommodate for any environmental settings
without manually updating the new node locations.
Localization is the technique that determines object positions from any
given distance or proximity information with respect to pre-defined reference
points. Prior to providing meaningful location information of an object, a global
reference has to be defined. For example, the most common reference used in
modern localization application is global positioning system (GPS) coordinates.
However, for low-cost localization systems, like [1], that do not have the luxury of
equipping GPS, it is common to define the location reference in 2-D Cartesian
space. Later in the related work section, we will discover that in order to localize
nodes in WSN, most localization techniques are required to have knowledge of
2
the inter-node distances. The objects‟ positions can then be computed with a
distance to position coordinate transformation algorithm, such as multi-
dimensional scaling (MDS) algorithm [2], which will be discussed later in this
thesis.
In WSN localization applications, obtaining inter-node distance information
is essential prior to performing node localization. Some of the most common
distance measuring techniques, described in [3], utilize ultra-sound or laser.
Range measurement with the above sources can be determined by the time a
signal travels from one node to another. This ranging technique is known as time
of arrival (TOA) and is also a commonly used technique in tracking applications,
such as Radar [4]. However, WSN localization systems utilizing TOA can be
costly and complex, due to additional source generation hardware and algorithm
to perform the processing. Also, TOA requires line of sight (LOS)
communications between nodes, making this technique limited to outdoor or
obstacle-free environments to provide accurate localization.
Without introducing additional ranging devices to localize WSN nodes,
ranging with RF signals are also widely researched and discussed. Some of the
common RF ranging techniques are time difference of arrival (TDOA), angle of
arrival (AOA), and received signal strength indicator (RSSI) ranging. TDOA
measures distance between two nodes by multiplying the difference of time
traveled from one node to another with the speed of light. This technique requires
clocks on both nodes to be synchronized and to be precise enough to measure
the subtle difference in time of travel. However, integrating an on-board clock or
3
timer with high precision can be very costly. AOA incorporates directional
antennas on nodes to detect the angle of the arriving signal, and it localizes
nodes based on the direction of RF signal propagation incident on the antennas.
Same as TDOA, AOA requires additional antennas to localize, and therefore not
only increasing the cost, but also increasing system complexity. The RSSI
ranging technique is the simplest ranging technique, in terms of system
complexity and cost, among all the techniques. RSSI ranging is done by
modeling the RF signal propagation over distance. In theory, the RF propagation
model demonstrates the monotonic degradation of RF signal strength over
distance. However, in practice, as many researches, [5] and [6] have discovered,
due to variety of electromagnetic (EM) interferences and reflection, RSSI values
do not always reflect its corresponding ground-truth distances.
The ranging techniques discussed above are known as the range based
measurement techniques. Techniques, known as the range-free based
measurement techniques, are also common implementations in WSN localization
systems. Most range-free localization is designed to provide proximity detection
based on hop counts. Ranging based on hop-counts only depends on
communication topology of the WSN and does not require additional ranging
devices. Range free localization systems are low in system complexity, but they
provide poor localization resolution.
Localization accuracy is highly dependent on range measurements. A fine
ranging technique results in high reliability in position computation. However,
every ranging technique exhibits limitations depending on the environment
4
settings and interferences. As a result, the objective of this thesis is to provide a
versatile, low-cost, low-power localization system. This thesis proposes a pure
RSSI based localization system that provides better localization performance and
a more cost-effective approach than the conventional RSSI based localization
methodology in situations where distance information is faulty and missing due to
disconnected communication between node pairs.
In later sections, we will discuss in detail the design implementation of our
localization system. Unlike many of the related works, we propose a localization
system that requires no additional hardware and utilizes only information
provided by RF signals to do both ranging and localization with reduced
computational complexity to accommodate for low-cost and low-power
applications.
1.2 The Self-Organizing and Self-Calibrating (SOSC) WSN
Accuracy of range measurement between each node pair within a WSN poses
an essential problem to the accuracy of node localization. To address the
instability issue of the RSSI, caused by EM interferences, the localization system
has to have the ability to reduce the influence of the faulty inter-node RSSI
values and to localize the whole network based on the relatively reliable RSSI
values. The solution to the above problem is called the SOSC WSN, which
localizes the network nodes with the following features
Real-time modeling of the RSSI to distance relationship; it is designed to
self-calibrate to accommodate for EM interference and noise
5
A self-organized system and network architecture that allows the nodes
within the WSN to communicate to retrieve inter-node distance data
MDS-based position computation and mapping algorithm that is used to
translate the inter-node RSSI values into their corresponding 2-D positions
The SOSC WSN consists of a range modeling unit (RMU), which in real-time,
computes parameters of a radio propagation model, which translates the inter-
node RSSI values into a distance estimate. The RMU provides a real-time
distance metric for RSSI in the environment to cope with the ever changing
random noise and interference. With the distance metric defined, the SOSC
system utilizes the novel MDS-based algorithm, proposed in this thesis, for
commutating and reconstructing the ground truth network layout. The above
components of SOSC WSN allow the nodes to localize themselves with simply a
radio. The radio, in this case, is a 2.4 GHz on-chip radio, provided by the Texas
Instrument (TI) CC2530 system-on-a-chip (SOC). This radio is designed to be
compatible with the ZigBee standards. This choice of hardware allows SOSC
WSN to be a ZigBee based ad-hoc network, where each node within the network
participates in routing and forwarding data for other nodes by both single hop and
multi-hop communications. The ZigBee network topology allows the SOSC
nodes to communicate and localize one another even in regions with partial
connectivity.
1.3 Scope and Outline of Thesis
The primary focus of this thesis is the design implementation of a low-cost, low-
power, and low system-complexity localization system based on a ZigBee ad-hoc
6
mesh network [7]. This thesis also discusses the merits and weaknesses of
several popular WSN localization systems. Due to the inaccuracy of RSSI
ranging measurement in nature, this thesis is not concerned with comparing the
localization accuracy with systems utilizing more sophisticated and refined
distance measurement techniques, such as ultra-sound or laser ranging. The
design approach of the localization and mapping algorithm is derived from RSSI
to distance relationship experimented and realized by several literatures.
Therefore, this thesis is not concerned with reproducing works of characterizing
RSSI to distance relationship; instead, it builds on results from the literature to
support the algorithm design and to demonstrate the validity of the approach with
an extensive experiment.
Chapter 2 provides brief background information on ZigBee standard and
its communication topology that allows the sensor nodes to obtain the distance
information from one another; the background also covers the model and
techniques necessary to translate the inter-node distance information and to
reconstruct the physical network layout respectively. Chapter 3 is a review of the
related work done on WSN localization. This chapter categorizes localization into
two categories, range-based and range-free localizations. This chapter discusses
the localization performance and implementation practicality among the above
various localization algorithms and systems. Chapter 4 covers the design of the
SOSC system, in terms of the network, system architecture. Chapter 5 discusses
the localization and mapping algorithm that allow the SOSC WSN to self-localize
with sufficient amount of accuracy. Chapter 6 describes the SOSC system‟s self-
7
localizing ability in real-life scenarios to demonstrate that this novel localization
approach is able to reduce the inaccuracy generated by RSSI ranging to achieve
an equally good accuracy that enables localization in partially connected network
with less computation complexity. Chapter 7 will conclude and discuss the future
work for this thesis.
8
Chapter 2 - Background information
The SOSC WSN is designed based on wireless mesh network topology. With the
use of novel MDS based algorithm along with the network clustering technique,
the SOSC WSN is capable of computing its node locations under various
environmental settings. The above localization technique proposed in this thesis
is called the clustering MDS localization algorithm. In this background information
overview, three major components underlying the design and implementation of
the SOSC WSN will be discussed, and they are
ZigBee mesh network topology and communication
Mapping technique
Distance measuring technique
2.1 ZigBee Communications Overview
ZigBee, [8], is a specification for low-cost and low-power network protocol based
on IEEE 802.15.4 personal area network (PAN) protocol standard. The low-cost
and low-power features make ZigBee to a suitable protocol for WSN applications,
requiring high density of wireless sensor coverage and long operating time on
sensor devices running on small batteries. A ZigBee network consists of exactly
one coordinator, several routers, and end devices. A ZigBee coordinator initiates
the network and defines the network PAN-ID, acting as a trust centre and
repository for security keys; a ZigBee router can be configured to run applications
and to also pass messages to other router devices within the same network; a
ZigBee end device can act as a sensing device that communicates with routers
9
or coordinator, however it has no routing capability. ZigBee‟s network
communication is based on a mesh network topology. With mesh networking,
ZigBee achieves high reliability in data communication. A typical ZigBee mesh
network is illustrated in the following figure
B
C
F
A
G
E
D
Gateway
Routers
Figure 2-1: ZigBee Mesh Networking
In a ZigBee mesh network, all the wireless sensors are able to communicate with
one another. Their messages can multi-hop and be routed through their
neighboring router nodes to reach the destination. In ordinary situations, ZigBee
routers in the network discover the “shortest path” to relay the message to its
destination, which in most cases, being the gateway or the devices that are
capable of storing and computing network information.
10
B
C
F
A
G
E
D
Gateway
Routers
Figure 2-2: ZigBee Mesh Networking Discovering Path
ZigBee, being a mesh network, is robust in terms of data communication
because messages can be routed through any available router within the
network. In cases where any of the routers loses connectivity to the network,
shown in Figure 2-2, the message will be detoured to the next available “shortest
path”, the path with the least communication cost, to reach the gateway. As a
result, communication within a ZigBee network can still allow messages to reach
the gateway in a partially connected network.
In a ZigBee network, all devices can communicate with other devices by
broadcasting or by uni-casting. Establishing successful communications
between node pairs requires the following addressing components
11
Table 2-1: ZigBee Network Addressing Components
Name Description
Channel ZigBee contains 16 channels in the physical RF spectrum
PAN ID The address of a network within a particular channel
NWK Address The address of a sensor node within the network
Application Endpoint The address of an application within the network
Cluster ID (CID) An Message identifier for exchanging information within an application
From an application‟s point of view, one may deploy different ZigBee networks at
various locations for separate monitoring or sensing purposes. To differentiate
one network from the other, network channels and PAN IDs are designated to
each network to allow independent network operation. Within each individual
network, each wireless node communicates with one another by application
endpoints, specifying the application which the information belongs to, and by
CID, specifying the action to execute within a particular application. The network
address (NWK address), identifies the destination address of the wireless node a
particular message is directed to.
Any ZigBee node within a network can perform two different addressing modes
to serve different application purposes and to exchange information with other
nodes. The two addressing modes are
Broadcast: A broadcast message is intended to be received by all the
wireless modules within the network. Broadcasting is initiated by any of
the nodes transmitting a message to its neighboring nodes, and the
message is repeated and forwarded to the next nearby node until every
node within the network receives the message.
12
R7
R1
R4
R9
Broadcast Node
R3
R2
R8
R6
R5
Figure 2-3: ZigBee Broadcasting Mode
In Figure 2-3, the red shaded circle is the transmission range of the broadcast
node. The multi-hop nature of ZigBee protocol allows the broadcast message to
reach devices outside of the range. The degree of message forwarding, however,
can be configured by controlling the number of hops allowed to cover the
network. This regional broadcasting scheme is useful in applications where
nodes are only interested in exchanging information with its neighboring nodes.
Uni-cast: Unlike broadcasting, uni-casting is intended for a device-to-
device direct communication. Uni-casting is an effective way of
communication, because this messaging scheme does not occupy
13
communication channel of every node within the network. Figure 2-4 is an
example of uni-casting, where the node performing uni-casting is to
transmit the message directly to the node, R5‟s address. R5, however, is
out of the transmission range of the uni-cast node, and thus requiring an
intermediate router node, R2 to forward the message.
R7
R1
R4
R9
Uni-cast Node
R3
R2
R8
R6
R5
Figure 2-4: ZigBee Uni-casting Mode
In the above case, only R2 is occupied for routing the message to R5. This
implies that the rest of the nodes are still able to communicate with one another
to exchange information without being interrupted.
14
2.2 MDS Algorithm
As described in [2], multidimensional scaling (MDS) is a method that transforms
the similarity or dissimilarity, represented in Euclidean distances among pairs of
objects, into lower dimensional multidimensional space. In our particular
application, the objects are the wireless sensor nodes, and the dissimilarity of
interest is the physical distance among every pair of nodes within a SOSC WSN.
One of main purposes of utilizing MDS algorithm for data analysis is to provide
visualization of object distribution in space. For example, one may want to
visualize the locations of individual cities in Vancouver, BC, on a 2-D map from
the given distances between those cities. Table 2-2 shows the table of
dissimilarity between each city pair in metro Vancouver, and the plot in Figure
2-5 represents the two-dimensional MDS spatial distribution between the cities.
Table 2-2: Inter-city Distance Matrix
City Vancouver Burnaby Coquitlam Richmond Delta
Vancouver 0 10 26 10 27
Burnaby 10 0 8 24 17
Coquitlam 26 8 0 23 15
Richmond 10 24 23 0 16
Delta 27 17 15 16 0
15
Figure 2-5: Vancouver City Layout
The objective of utilizing MDS is to transform the complex dissimilarity into a
simple visual layout, like Figure 2-5. In two-dimensional space and to also
represent object locations in Cartesian coordinates which can be universally
converted into other coordinate standards, such as GPS coordinates. To
understand the underlying theory of MDS, the following terminologies have to be
defined
2.2.1 Dissimilarity Matrix
As mentioned earlier, dissimilarity or similarity of any pairs of objects indicates
their likelihood. This likelihood is expressed in forms of distance metric. In
classical MDS (CMDS), this distance is calculated with Euclidian distance
formula
(1)
16
Where dij is the inter-object distance between ith and jth object x1,…,xi . The
dissimilarity matrix is the table of all the inter-object distances, calculated for all
pairs of i,j I . The resulting dissimilarity has the form
(2)
Having the dissimilarity matrix, the question remains how to transform these
inter-object distances into a coordinate standard to provide a visual
representation of the objects‟ corresponding spatial distribution. A common
method to solve the above problem is the following optimization
(3)
The solution can be found by performing Eigen-decompositions on the matrix.
2.2.2 Eigen-decomposition for MDS
Eigen-decomposition is a technique in linear algebra to factorize a matrix into its
Eigen vector and Eigen values. For a NxN square matrix A, using Eigen-
decomposition, can be expressed as
(4)
Where v and λ are the Eigen vector and Eigen value of matrix A respectively.
Error! Reference source not found. can be rewritten with Eigen vector qi, for (i
1,….,N). The new expression is shown below
(5)
17
In MDS algorithm, matrix A is the dissimilarity matrix, whose distances are
calculated from the set of vector of coordinates, X. To perform the Eigen-
decomposition for dissimilarity matrix, there are several steps to follow,
Calculate the squared dissimilarity matrix: The square of a dissimilarity
matrix can be expressed as D(2), where
(6)
The vector c is the vector containing the diagonal matrix elements of XX’,
which can be expressed as vector B, also referred to as the matrix of scalar
products.
Apply double centering to the squared dissimilarity matrix: Double
centering is a method to subtract row and column means of a matrix from
its elements and to add back the grand mean. After applying double
centering by multiplying the centering matrix J on both side of matrix D(2)
and multiplying the whole vector by a factor of
, the squared
dissimilarity matrix D(2) becomes,
(7)
Calculate the Eigen-decomposition for matrix B: To compute the MDS
coordinates of matrix B, the same factorization method stated in (5) is
18
applied to obtain the corresponding Eigen vectors and values. Error!
eference source not found. can be rewritten as follows
(8)
The final solution is determined by multiplying the first m positive Eigen
values Λ+ with the first m columns of the Eigen vectors Q+. The MDS
coordinates, X, becomes
(9)
The above algorithm describes the procedure to factorize the inter-object
distance matrix into their corresponding Eigen vectors and values. This solution
is the best-fit, however, not unique, and it consists of rotational and translational
offsets from the actual objects‟ spatial layout. The following section describes the
procedure to transform the best-fit MDS solution into the solution that matches
the actual inter-object layout in space.
2.3 Iterative Closest Point (ICP) Algorithm
MDS algorithm defines the best-fit solution only with respect to the dissimilarities
among nodes within a network. However, visually speaking, the best-fit solution
is meant to be fitted to the global ground truth reference. In real-world scenario,
this ground truth reference is referred to as anchor points. The best-fit solution is
produced by determining the optimal positions to register the MDS-computed
node positions to match with the anchor points. This mapping optimization
19
procedure can be achieved by applying the ICP algorithm, developed by authors
in [9]. ICP is performed by minimizing the following equation
(10)
Where are rotation and translation required to be done to register the
nodes into the most optimal positions. is the weight, equal to 1, if is the
closest point to , otherwise . This algorithm iteratively searches through
the number of points in the model set until the closest points are found. The
final result produces the rotational and the translation matrix, whose theoretical
backgrounds are described in the following section.
2.3.1 Rotational Matrix
As mentioned previously, MDS algorithm provides the best-fit coordinates, X,
based on the dissimilarity matrix B. This best-fit solution, however, can contain
rotational or translational ambiguity in space.
Figure 2-6: Rotational and Translational MDS Solution
The plot on the right in Figure 2-6 is the example of a best-fit solution computed
by MDS algorithm. The plot on the left is the actual node layout in physical
20
space. It can be shown that after performing MDS on the dissimilarity matrix from
the actual layout, the best-fit solution is rotated and translated. To obtain the
actual layout from the solution provided by MDS algorithm, two or three anchor
points have to be pre-defined. Anchor points provide fixed coordinates that limit
the degree of freedom for the MDS solutions, so its solution can be correctly
oriented to fit the actual object layout in physical space. A rotational matrix, R, is
generally defined as
(11)
The above equation can rotate a set of object coordinates by an angle θ in the
clock-wise direction. To perform rotation in the counter clock-wise direction, (11)
becomes
(12)
The angle θ is defined by the angle separation between the two anchor points
and their corresponding MDS solutions. The angle can be calculated from the
following equation
(13)
Where vectors a and b are defined by the actual two anchor positions and their
corresponding MDS solutions respectively. To calculate the angle of separation,
the MDS solution in Figure 2-6 is first translated to overlap the actual plot at the
origin, defined by point A. The new plot becomes
21
Figure 2-7: Translated MDS Solution and its Angle separation with the Actual Layout
From (13), the angle of separation between the actual layout and the MDS
solutions is calculated, and its value is expressed in the range, where
, in either the clock-wise or the counter clock-wise direction. To rotate the
whole set of objects by angle, , in vector space, the rotated object coordinates
are expressed as follows
(14)
2.4 Radio Frequency Ranging Techniques
In literatures, WSN localization applications are usually used to compare with
GPS localization, because of their low-power and low-cost advantages over GPS
localization. Bulusu in [1] describes the design of low-cost GPS-less localization
system with the use of only RF communication capabilities. As previously
θ
22
discussed, most localization methods require distance or proximity information
between nodes to compute node positions. This inter-node distances can be
determined by various methods, such as ultra-sound or laser. However, ultra-
sound waves and laser are easily absorbed and reflected by obstacles. As a
result, ultra-sound and laser ranging techniques become limited to perfect line-of-
sight (LOS) environments.
To reduce the system cost and complexity and to be able to localize in a
non-LOS environment, RF wave is also often used for measuring pair-wise
distances between radio devices, [10]. The most common RF ranging techniques
are
2.4.1 Time Difference of Arrival (TDOA)
According to Xu, [11], TDOA is a commonly used ranging technique. In the
context of range measurements implemented by Xu, UWB radio technology is
used. Radio waves are forms of electromagnetic radiation, traveling at the speed
of light, . The inter-node distances between a pair of radio
devices can be calculated from a radio signal‟s traveling time from one device to
the other. The distance traveled between radio devices a and b can be
approximated by the following equation
(15)
Tab is the time taken by the radio signal to travel from device a to device b. To
measure this time difference, the clocks on both radio devices are required to be
synchronized. Any mismatch in the timer synchronization on both sides will result
23
in imprecision in distance approximation. The following illustration is the typical
wireless sensor localization network based on TDOA.
Mobile User
Anchor 1
(X1, Y1)
Anchor 3
(X3, Y3)
Anchor 2
(X2, Y2)
Figure 2-8: TDOA Triangulation based Localization
Figure 2-9: Time Traveled between Mobile Node and its Neighbouring Anchors
dM1
dM3
dM2
Time
TM1
TM2
TM3
24
Figure 2-9 demonstrates the time taken for mobile node‟s radio signal to reach its
three neighboring anchor nodes. Using (15), the corresponding distances of the
time difference of RF signal arrival can be computed.
2.4.2 Receiver Signal Strength Indicator (RSSI)
Another common ranging method is utilizing the RF signal propagation
characteristics. In a typical radio communication system shown in Figure 2-10,
the receiver signal strength, Pr, is modeled by Friis equation, [12], in (16)
Transmitter Reciever
Figure 2-10: Basic Radio Communication System
(16)
Where Pt is the transmission power; Gt and Gr are the transmitter and receiver
antenna gains respectively; λ is the wavelength of the RF signal; R is the
distance between the two radio devices. It is obvious that the receiver power is
mainly governed by its distance from the transmitter and its power decays in an
inversely proportional relationship with the squared distance. The characteristic
curve is shown in Figure 2-11. When both the transmitter power and receiver
power are known, the distance between the two radio devices can be
approximated.
25
Figure 2-11: RSSI vs. Distance Characteristic Curve
However, it is difficult to observe accurate receiver and transmitter signal
strength without a proper measuring device. In modern radio receivers, their on-
board controllers are able to detect the power present in the received signal. This
measurement of received signal power is referred to as RSSI. According to
Texas Instruments [10], the relationship between RSSI and distance can be
modeled by the following equation
(17)
Where n is the propagation constant; d is the distance; A is defined as the
received signal strength at distance of one meter.
26
Chapter 3 - Related Work
As discussed in Chapter 2 -, to develop a SOSC WSN, the building blocks
required are, a wireless network capable of performing single hop and multi-hop
communication between the sensor nodes; a robust sensor position computation
technique; a reliable distance measuring mechanism to maintain the accuracy
and consistency of the localization result. This section describes the related work
utilizing various distance measuring techniques, network topologies, and
algorithms to perform localization in WSNs. Unlike [3], treating the ranging
techniques, position computations, localization algorithms as separate entities of
localization systems, all the related works in this section are categorized into two
main groups, range based and range-free localization systems
3.1 Range Based Localization Systems
Range based localization systems take advantage of distance measuring devices
to obtain an overall distance measure between a pair of nodes within a WSN.
With the use of various localization algorithms, the positions of the sensor nodes
can be computed from the inter-node distances. The following are examples of
related works on range based localization systems. The authors of, [13], [14],
and [15], introduce modern range-based localization techniques and
implementations for WSNs. Various measurement techniques mentioned in their
works, such as TDOA, RSSI profiling, GPS ranging, and AOA [16], are common
approaches to obtain inter-node distance prior to performing localization. Given
the inter-node distance, range based localization systems utilize stochastic
27
methods or geometric triangulation to perform the conversion from inter-node
distances to their corresponding Euclidean coordinates. The following examples
describe the details of various implementations
3.1.1 RSSI Based Localization Systems
RSSI based localization systems are the simplest and cheapest, because they
utilize only radios to perform distance estimation and do not require additional
distance measuring devices that increase the overall complexity of individual
sensor devices. The localization technique using RSSI was first adapted in
mobile networks in [17]. This author proposed a weighted least squared (WLS)
method to compute the positions of mobile stations (MS), having true location of
, from their distances to each base station (BS) with coordinates
. The RSS to distance relationship is given by the model,
, where
is the transmission power in dBm, is the
factor influencing the signal strength, and is the propagation constant. The
RSS, denoted by , can be represented in 2-D space by the following equation,
. The WLS method
essentially solves the system of linear equations of each MS to BS pair distance;
the system of equations can be generalized by the matrix form, , where
,
, and
. With WLS method,
the localization simulation results have shown the mean square range errors
(MSREs) lower than the standard least squared (LS) method. The LS approach
28
is also known as a dimensionality reduction method. Some similar approaches
based on dimensionality reduction are introduced in [18], [19], [20], and [21]
where the most common localization approaches use the MDS algorithm. In [21],
Aloor and Jacob introduce non-MDS based algorithm, called the stochastic
proximity embedding (SPE), to provide a comparable localization accuracy with
less computation complexity than the classical MDS methods. As concluded by
Aloor and Jacob, the competitive advantage of SPE over other MDS-variants is
its comparable localization performance, lower computation complexity, and its
ability to localize in irregular network topologies; however, being highly
dependent on distance information, the localization accuracy of the above
algorithms are all subject to error in noisy environments. This thesis proposes a
MDS-based localization approach, which not only has a linear scaling property in
complexity, but also has the capability to reduce the influence of range error to
produce lower localization error than the classical MDS method.
Range measurement is a major component in range-based localization.
As for RSSI-based ranging, mathematical models for converting RSSI to distance
follow the inverse squared relationship, as illustrated earlier in Figure 2-11.
However, as observed in [22], interferences in the RF channel can cause the
inter-node distance to enlarge, causing high localization error. Many works have
been done to model various RF channel interferences. Depending on the
particular RF interference the equation attempts to model, the complexity of the
equations may vary. For example, in [23], the RSSI to distance relationship in a
RF shadowing model is expressed by the equation,
29
. The path loss exponent, η, as experimentally determined, can
vary between 2 and 6, depending on the terrain and the surroundings in pair-wise
LOS communication. However, for simplicity, not considering the RF channel
fading, Wu utilizes (17) to obtain distance information. To compute the
corresponding positions from the inter-node distances, Wu proposed mean
filtering to localize nodes. Mean filtering is a low pass filter that reduces the
influence of impulse noises or sudden change in RSS, smoothing the signal. In
the 12x12 m2 indoor environment, this particular localization system introduces
an average error of 1.6 meters.
The previous two works perform localization based on distance extracted
from theoretical models of RF wave propagation. A famous research, RADAR
[24], done by Bahl and Padmanabhan, proposed an empirical approach by
constructing a map of overlapping signal strength coverage of the BSs that
define the boundaries of the WSN. The map is determined by recording the
signal strength information in various locations within the network. As a result,
each location will present a different combination of signal strengths transmitted
by the neighbouring BSs. From comparing the empirical method with the
theoretical method, using the propagation models, the author claimed that
RADAR system can achieve a median resolution of 2 to 3 meters, as opposed to
a median resolution of 4 to 5 meters in the theoretical methods.
The empirical localization method is always done off-line, as the system is
required to have knowledge of the environment prior to tracking sensor nodes
within the areas of interest. For the localization system, requiring such an off-line
30
training, [25], [26], and [27] proposed stochastic methods to design rule-based
localization algorithms that determines the most probable node positions from the
RF signal strength information gathered. The stochastic localization algorithms
from the above three references utilizes methods like maximum likelihood
estimator (MLE), maximum a posteriori probability (MAP), and the hidden Markov
model (HMM). One characteristic that all stochastic localization systems have in
common is that all systems can be trained to have high accuracy and high
reliability in noisy environments. However, they require extensive off-line
environment data collection and thus are not suitable for applications requiring
rapid network deployment.
3.1.2 Non-RSSI Range Based Localization Systems
Although, RSSI plays an important role as a ranging method in localization
systems, many works have been done on using other techniques to obtain
distance measurements. Those techniques can be categorized as follows
TOA/TDOA Based Localization: As previously discussed, radio signal
strength is a commonly used ranging technique in WSN localization
applications. Time can also be used to interpret distance between two
nodes in a WSN. TOA and TDOA are widely applied time-based ranging
and tracking techniques. TOA requires all transmitters to be synchronized
with the anchors, so the travelling time from a transmitter signal to each
anchor will determine transmitter-to-anchor distances, which allow the
transmitter‟s position to be trilaterated. TDOA, on the other hand, only
concerns with the difference between arrival times from one transmitter to
31
the anchors. The advantage of TDOA is that it requires only the anchors to
be synchronized. Common WSNs have more transmitters than anchors.
Using TDOA localization can greatly reduces system complexity and
network overhead for transmitter-to-anchor synchronizations. Many
variants of the TDOA based localization algorithms, like [28], [29], and
[30], are introduced to deliver better accuracy and reliability under noisy
environment. However, the above localization algorithms rely on solving
the systems of squared distance,
,
where is the TDOA measured distance from different reference nodes
to the target node. Although, TDOA requires much less synchronization
than TOA localization. Synchronization is still a difficult procedure, for it
requires highly precise clocks and sophisticated inter-anchor
communication scheme to synchronize their clocks. Besides RF waves,
sound waves are also common ranging sources in TDOA localization
systems, such as the Cricket localization system, [31] and [32]. TDOA
using ultrasound sources does not require clocks to be as precise as
those used for synchronizing RF waves. However, TDOA methods will
always induce network overhead for synchronization and thus is not
suitable for low-cost and low-power applications, whose network
bandwidth and device clock precision are limited.
Angle of Arrival (AOA) Based Localization System: a standard AOA
method is initially designed to determine the direction of the signal
propagation. The receivers of an AOA based localization systems are
32
often equipped with antenna arrays. The angle of the arriving signal is
determined by calculating the time delay of signal at each antenna. In [33],
the author used a probabilistic approach to determine the position of
nodes within the network by using only the AOA information. This is done
by first assuming all the transmission between nodes are bounded by a
maximum transmission range of , and any information beyond the
transmission range will be neglected. Second, all the nodes are assumed
to be in LOS communication with one another without any obstruction
within the network. Having defined the above two assumptions, the AOA
information received by a target node , is modeled with a Gaussian
distribution,
. Translating the AOA information to
the corresponding Cartesian coordinates, the author proposes using joint
pdf between distance, and angle, . The Cartesian
coordinate can be realized by transforming this joint pdf. The transformed
pdf is
Where is a polar angular function of and :
By calculating the probability distribution intersecting at unknown node
from its neighbouring anchors, the position can be determined. Other AOA
33
based localization systems such as [34] and [16] proposed triangulation
algorithms to demonstrate comparable localization accuracy.
3.1.3 Hybrid Range Based Localization System
Being fading and interference-prone, RSSI ranging can be unreliable and
erroneous. To avoid compromising the localization accuracy due to RSSI range
measurement error, two approaches can be applied. One approach is to train the
system to determine the trust-worthiness based on prior knowledge of range
measurement distribution, where higher weights are assigned to the
measurements believed to be more accurate. The second approach is to utilize a
more reliable ranging device, which is less prone to noise than RSSI. Like the
Cricket system, ultra-sound is used in conjunction with RSSI to compute inter-
node distances. In [35], the author combines the above two approaches to
address the ranging error issue caused by RSSI. It is important to remark that
this literature has the most similar approach as the SOSC system proposed in
this thesis and will be used as comparison throughout the thesis.
The algorithm used in [35] is called distributed weighted MDS (dwMDS)
algorithm. This algorithm is designed to operate with a broad range of ranging
methods, such as RSSI or TOA. Based on the range measurements obtained, a
cost function is defined as
Where
is the weight, equal to 0 when the measurement
is not available
between and . when the measurement is believed to be more
34
accurate. However, this literature does not elaborate on the methodology for
differentiating an accurate range measurement from many other measurements.
The above assumption has allowed the algorithm to be non-parametric, but its
adaptive nature for searching for the more accurate range measurement results
in a computation complexity.
3.2 Range-free Localization Systems
Range based localization algorithms can achieve high accuracy in position
estimation. Due to additional ranging devices required to provide distance
information, the complexity and cost of the overall system can be high. Range-
free localization systems do not require metric distance information to perform
location calculation. Range-free localization utilizes connectivity information, such
as, “who is within the communications range of whom” [36], to approximate the
position of nodes. As a result, the systems are simpler and more cost effective
while sacrificing localization accuracy. Many of the range-free localization
algorithms are developed based on connectivity information as discussed by [37].
The most popular range-free localization algorithm, utilizing the distance vector
(DV) hop approach, is proposed by [38] and is further modified and improved by
[39]. There are also other range-free approaches like the approximate point-in-
triangulation test (APIT) approach in [40], and the proximity distance map (PDM)
algorithm in [41]. These algorithms will be discussed in the following sections.
3.2.1 DV-Hop Localization Algorithm:
The first DV-hop algorithm was proposed by Niculescu and Nath [38]
specifically for ad-hoc networks. An ad-hoc network, by nature, transmits
35
information throughout the whole network through hopping. Such an ad-
hoc based localization system is also known as an ad-hoc positioning
system (APS). APS performs localization based on the hop-by-hop
characteristic in a distributed fashion and is an effective localization
system when the network coverage is large, where not every node is
within the communication range of every other node. In a DV-hop based
APS, the network is bound by a few anchor nodes with known inter-node
distances from one another. These anchor nodes are known as the
landmarks , where . The proximity information is governed by
the correction function, , where
Each landmark node has its correction function. The function is defined by
the ratio of inter-node distances, represented with two-dimensional
Cartesian coordinates, and , and the sum of hops, . The localization
algorithm based on DV-hops uses triangulation. This particular APS has
the knowledge of the number of hops that each node is required to reach
each landmark node. The proximity information is obtained by multiplying
the landmark-specific correction function to the number of hops the target
node is required to reach it. The proximity information of the landmark
nodes to the unknown node is triangulated, and its position is then
estimated. One advantage of DV-hop localization is that the localization
result is not affected by distance measurement error. However, in real-life
36
scenario, hop sizes from node to node may differ, and thus reducing the
localization resolution. To refine the accuracy of the localization results,
the author proposed a method, called DV-distance, where instead of using
hop counts as the proximity measurement, the localization system uses
RF signal strength. This method is also proposed in [42]. However, this
method, although is theoretically more accurate than the regular DV-hop
localization algorithm, in reality, raw RSSI values do not always reflect the
actual distances, due to RF interference and noise. Unless, a careful
RSSI-distance model is considered, the assumption of distance directly
based on RSSI can be erroneous in real-world environments. The author
in [43] mentions that ordinary DV-hop based localization‟s accuracy
suffers from the hop-count shift problem, which is mainly caused by the
assumption of ideal radio communication models. In this paper, the author
proposed an improvement to the ordinary DV-hop localization by
introducing a more complex localization mechanism that considers various
radio channel behavior, such as shadowing.
3.2.2 The Approximate Point-In-triangulation Test (APIT) Approach
Although, the DV-hop localization technique is a simple and an effective
method to provide a rough location estimate for nodes in large area
networks, the technique is sensitive to localization error when the network
is not isotropic. In real-life scenario, radio patterns and node placements
are usually random due to the nature of radio propagation and the
presence of obstacles, disrupting the LOS between any pairs of nodes
37
within a network. An APIT based localization system solves the irregular
node placement problem as discussed above. An APIT localization
algorithm utilizes anchor nodes to form an enclosed region to determine if
a target node is within or outside of the centroid, defined by the enclosed
region. The enclosed region is a triangle, formed by any of the three
anchor nodes within the network. In any APIT localization system with
anchors, there will exist triangular regions. The APIT algorithm
computes the set of centroids for the triangles, namely the centre of
gravity (COG) [40], where the target node belongs to. The algorithm
searches through all the triangles to locate and to finalize the target node‟s
position when an accuracy threshold is reached. APIT algorithm, although,
does not rely on the conversion of RF signal strength to distance to
perform localization, it relies on RF signal strengths between the target
node and its neighboring anchors to determine whether the particular
target node belongs to the centroid of the corresponding triangle. As a
consequence, the localization result can be faulty if any of the anchor
node reports incorrect anchor-to-target signal strength, positioning the
target node in an incorrect location. However, the author has proven that
APIT has comparable accuracy and a reasonably small computation
overhead against other range-free localization schemes, such as the DV-
hop localization algorithm.
38
3.2.3 The PDM Localization:
As mentioned earlier, the anisotropic characteristic of a typical WSN is
mainly caused by irregular terrain of the network, anisotropic RF radiation
pattern, and a random placement of the sensor nodes. This characteristic
greatly degrades the accuracy and reliability of the range-free hop-based
localization algorithms, such as the DV-hop algorithm discussed previously.
To address the anisotropic nature that pure hop-based localization
schemes are unable to solve, [41] proposed applying the transformation of
the proximity measurement between sensor nodes into a geographic
distance embedding space by utilizing the truncated singular value
decomposition based (TSVD-based) technique, called PDM. PDM is
defined by a linear transformation , consisting of two major components:
the M-by-M proximity matrix and the geographic distance
matrix . The linear transformation is represented by
. To estimate the geographic distance , the author applied the
singular value decomposition (SVD) to derive a truncated pseudo-inverse of
, . PDM can reconstruct the embedding space for geographic distance
using the proximity measurements. The estimate of the geographic
distances is , where is the proximity vector obtained
from the hop count to the anchor nodes. This PDM based localization
method‟s performance is compared with methods like the DV-hop and
MDS-MAP method and is shown to outperform the two methods both under
isotropic and anisotropic conditions.
39
3.3 Chapter Summary
The objective to perform localization in a WSN is to allow the sensor nodes to
realize their geographical positions among themselves without user intervention.
The most essential component for position computation is the distance
information between each pair of nodes within the network. This distance
information can be expressed in form of absolute distance (range-based) or
proximity (range-free). Based on these two types of distance information, various
localization algorithms have been developed. However, each type of localization
algorithm exhibits its weakness. For range-based localization systems relying
heavily on prior knowledge of inter-node distance distribution, off-line training and
data collection are required. Although, their localization performance is highly
reliable under noisy conditions, it is inefficient and inconvenient having to train
the system for different environment settings. Systems that do not require off-line
training rely mostly on linear transformation to translate the inter-node distance
information into corresponding coordinates. In general, range-based localization
has a finer resolution than range-free localization. However, ranging techniques,
using radio or ultra-sound, are prone to error in environments with reflective
media and multi-path effects. As a result, the localization algorithms often
incorporate statistical approaches to cope with these conditions.
Range-free localization schemes, on the other hand, are more reliable,
due to a low requirement in resolution demanded by its applications; it is often
sufficient to localize sensor nodes in terms of proximity, instead of the absolute
position by coordinates. However, range-free localization systems‟ performance
40
can be affected by their network topology and network characteristics. For
example, most of the range-free localization algorithms perform computation
based on hop-counts between nodes and the anchors. The network topology and
the degree of isotropy define the routing table, which also defines the hop range.
In an anisotropic network, the physical distance, requiring one hop to
communicate between two nodes, can require more than one hop for other
nodes, depending on the communication cost. This disadvantage is
demonstrated by the DV-hop localization scheme. However, it can be improved
and made more robust in anisotropic networks by increasing the density of
anchor nodes or by utilizing the APIT and PDM algorithms.
41
Chapter 4 - SOSC System Architecture
SOSC is a ZigBee based WSN localization system, whose network topology is
inherited from the ZigBee topology. The SOSC WSN is intended for sensing and
monitoring applications, and self-localizing capability is the feature that facilitates
the deployment of the WSN. SOSC WSN consists of only one gateway, which is
also a ZigBee coordinator, numerous routers and end devices. In this section we
describe the SOSC system architecture, the network, hardware, and also the
firmware architecture.
4.1 System / Network Architecture
GatewayPC App
Neighbor_A Neighbor_B
Neighbor_C Neighbor_D
Report_Node
1: Status
Report
2: RSSI
Request
3: RSSI
Forward
Figure 4-1: System and Network Architecture
42
In the context of this thesis, we will focus on the localization component of the
SOSC system. Prior to performing localization, the most essential component
required for computation is the inter-node distance information. In a SOSC WSN,
Figure 4-1, the gateway is responsible for receiving the inter-node distance from
the devices to be localized. The gateway, interfaced to a PC, will forward the
inter-node RSSIs through serial interface for further position computation. The
gateway is only responsible for receiving inter-node distances and plays no role
in the localization of the SOSC WSN. On the other hand, the routers and end
devices are the ones responsible for establishing the communication among one
another to report the inter-node RSSI to the gateway in an orderly fashion.
The reporting of inter-node RSSI values is initiated by a periodic and
asynchronous node status report. This status report denoted by „1‟ in Figure 4-1,
developed by a former CiBER colleague, Benny Hung [44], is used to reduce the
congestion of the channel traffic in a large scale network. Each status report
consists of the device ID of the reporting node and its current status, such as the
on-board battery level or the temperature in the surroundings. Only one node is
allowed to report status at each allocated time slot. Upon each status report, the
status reporting node broadcasts a request, denoted by „2‟, to its neighboring
nodes within one hop distance to forward the RSSI, denoted by „3‟, received from
the reporting node to the gateway. The inter-node RSSI forwarded to the
gateway is a packet that contains the device ID of the reporting node, the device
ID of the node receiving the request, and the RSSI value received.
43
The computing unit, in this case, is a PC, interfaced to a serial port of the
gateway device, obtains the inter-node RSSI packets forwarded directly from the
gateway upon reception. The PC is responsible for storing and organizing the
inter-node RSSIs into a square dissimilarity matrix with its RSSI values indexed
with the reporting node‟s and the receiving node‟s device IDs. The PC computes
the nodes‟ positions with a centralized MDS-based algorithm. This algorithm
computes and reconstructs the map of the node distribution in 2-D space from
the dissimilarity matrix. The PC also provides a visual realization of the node
layout in a graphical user interface (GUI), which allows the users to easily
understand the spatial distribution of the nodes within the SOSC WSN.
4.2 SOSC System Hardware Architecture
CC2530
Microcontroller
(8051 Core)
12-bit
ADC
2.4 GHz
RadioUART
CPU/RF Unit
Sensors
RS232
ConnectorPC
Computation Unit
Sensory Unit
Figure 4-2: Hardware Architecture of a Gateway Device
44
It is important to note that the underlying hardware devices of the SOSC WSN
are engineered by other CiBER researchers. Since, this thesis only concerns the
localization feature of the SOSC devices, the discussion on hardware
architecture will only focus on device power consumption, data collection
interfacing, and communication.
The hardware architecture of the gateway, routers, and end devices in a
SOSC WSN share similar design and are all based on the TI CC2530 system on
a chip (SOC). Figure 4-2 is an illustration of the major SOSC system building
blocks. The CPU/RF unit governs the inter-node communication and sensing of
the inter-node RSSI. The sensory unit has a 14-bit ADC on the CC2530 SOC.
When equipped with variety of sensors, it is capable of sensing the surrounding
temperature, on-board battery level, or vibration voltage. This unit provides
necessary information on each device‟s operating status upon each status report.
The computation unit is interfaced directly to the on-chip UART port. Through this
serial connection, a PC is able to obtain the status information of individual node
and the inter-node RSSIs with its neighboring nodes. Depending on the type of
devices and the role it plays in the SOSC WSN, SOSC nodes do not require all
functional units. The table below shows the functional units required for each
type of devices.
Table 4-1: Table of Required Functional Units on each Type of Device
CPU/RF Unit Sensory Unit Computation Unit
Gateway/PC Yes No Yes
Router Yes Yes No
End Device Yes Yes No
45
The gateway‟s major function is to gather and forward the network information to
the PC; therefore, it does not require additional hardware on-board. On the other
hand, router and end devices are the sensing and tracking units within the WSN.
Therefore, routers and end devices are often equipped with sensors, capable of
conducting environmental or physiological monitoring. Figure 4-3, Figure 4-4, and
Figure 4-5 are the actual ZigBee based wireless nodes used in the SOSC WSN.
Figure 4-3: SOSC Gateway Device
CC2530 Motherboard
RS232 Connector
46
Figure 4-4: SOSC Router Device
Figure 4-5: SOSC End Device
The gateway device is configured on a CC2530 evaluation board (EB) in Figure
4-3, which has the on-board RS232 connector interface and can easily be
programmed to receive data from nodes in the WSN; the data is communicated
directly with a PC through UART. The routers and end devices are designed to
Power/Debug Socket
CC2530 Motherboard
CC2530
Power/Debug Socket
47
consume minimal power and therefore few I/O peripherals and a limited number
of application-specific sensors are required. For the localization purpose in this
thesis, we utilize the on-board components listed below
Microcontroller: SOSC nodes use TI‟s CC2530, consisting of an IEEE
802.15.4-compliant 2.4 GHz on-chip radio transceiver. During the on-
time, the radio component consumes 28.7mA in Tx mode and 20.5mA of
current in the Rx mode respectively. For serial data communication,
CC2530 contains universal synchronous/asynchronous
receiver/transmitter (USART), which can be configured as either serial
peripheral interface (SPI) or as an UART. CC2530 is a low power
microcontroller that consumes 3.4mA of current under normal CPU
operation without radio activity and 1μA under power saving mode.
Debug/Power Connector: the debug and power connector on SOSC
routers and end devices serve two purposes. The most important purpose
is to power the devices and the second purpose is to provide an interface
for programming the on-board microcontroller. This connector could also
be customized to provide interface to external sensors connected to the
SOSC devices, such as intrusion sensors for perimeter surveillance
applications.
RS232 Interface: the RS232 interface provides a direct data link to
external host devices such as a PC. Among the three types of SOSC
devices, only the gateway consists of this interface and is allowed to
exchange information with a PC directly.
48
Antenna: antenna provides gain for the radio signal to propagate afar,
exchanging data and conducting localization among all the SOSC nodes.
SOSC WSN utilizes omni-directional antennas to communicate and
collect inter-node RSSI information.
Figure 4-6: SOSC Antenna Radiation Pattern [45]
The localization approach proposed in this thesis converts RSSI into
distance, based on the radio propagation model discussed in (17).
Antenna property is a major influence on the signal strength polar
distribution in the antenna‟s LOS communication with others. This property
is realized with the radiation pattern, which provides the visualization of
directionality of the antenna gain. The antenna used for SOSC WSN is a
2.4 GHz omni-directional antenna from Antenova Ltd. It has a radiation
pattern as illustrated in Figure 4-6 and demonstrates equivalent antenna
gain in all directions in the XY plane; the pattern is less omni-directional in
the other two planes, and yet it still provides similar gains in all directions.
49
4.3 SOSC Software Architecture
The SOSC software consists of two major components: the first one is the on-
chip firmware, which provides functions for the gateway to receive and to
interpret inter-node RSSI packets and for the routers and end devices to request
the inter-node RSSIs from their neighbors on an asynchronous scheduling
mechanism; the second component is the software GUI on a client PC, which is
responsible for computing the localization algorithm based on the inter-node
RSSI and for displaying the estimated positions.
4.3.1 SOSC Firmware:
Hardware Drivers
OS/Kernel
ZigBee Stack
User Application
802.15.4
MAC
Figure 4-7: Firmware Architecture for SOSC Devices
Figure 4-7 shows the firmware architecture functional blocks for designing the
application running on the SOSC devices. Besides the user application layer, the
lower layers are proprietary and are provided by TI‟s Z-Stack. The OS kernel is
50
responsible for managing event loops and interrupts. The ZigBee stack and the
802.15.4 MAC layer govern the ZigBee protocol and the radio transmission
respectively. It is important to note that all the layers in the Z-Stack are
implemented in C.
The user application contains the processes allowing each node to
communicate with the rest of the network to gather the inter-node RSSI. As
mentioned previously, the node reporting the status initiates the request for RSSI
from its neighboring nodes. The initiation of the request and its interaction with
other layers prior to broadcasting is as follows
Figure 4-8: RSSI Request MSC between Z-Stack Layers
The status report is generated with a timer interrupt in each node in an
asynchronous and periodic manner. This interrupt is handled by the OS
abstraction layer (OSAL). Upon each timer interrupt, OSAL requests the
application layer to initiate the RSSI request to the neighboring nodes via
broadcasting, handled by the application framework interface layer (AF layer).
The broadcast message to the neighboring nodes has the following format
51
DESTINATION_ADDR CID REPORT_NODE_ID BROADCAST_RADIUS
2 Bytes 2 Bytes 2 Bytes 1 Byte
Figure 4-9: RSSI Request Broadcast Packet
The destination address indicates where the message should be directed to. This
address is often referred to as the IEEE 16-bit short address. For uni-casting, this
address is required to contain the address of the target device. For broadcasting,
this address should be coded with the predefined value by the Z-Stack. The
cluster ID, as specified in Table 2-1, is an ID that specifies the exchange of
information within an application. This CID for broadcasting is named
“PEG_ID_CID” and is to be identified once the over the air (OTA) broadcast is
received by the neighboring nodes. The broadcast radius indicates the number of
hops allowing the OTA message to transmit to reach its end-point. For RSSI
request broadcasts, the node performing the status report only requests the
RSSI values from its one-hop neighbors.
The message interaction between Z-Stack layers upon reception of the
RSSI request broadcast is illustrated in Figure 4-10
Figure 4-10: Receiving Node MSC between Z-Stack Layers
52
The OSAL layer acknowledges the application layer of the packet arrival, and the
application layer unpacks the message, whose content is shown previously in
Figure 4-9, where the CID acknowledges the application layer to obtain the RSSI
value from the broadcasted message and to create and to forward to the
gateway the new packet, containing the destination address, broadcaster‟s ID, its
RSSI, and the receiver‟s ID. The new packet has the following format
DESTINATION_ADDR
CID REPORT_NODE_ID RSSI RECEIVER_NODE_ID BROADCAST_RADIUS
2 Bytes 2 Bytes 2 Bytes 1 Byte
2 Bytes 1 Byte
Figure 4-11: RSSI Forward Packet
This packet containing the RSSI information is directed to the gateway via a
multi-hop uni-casting mechanism. The packet specifies the address of the
ZigBee coordinator, gateway, to be always 0x0000. The CID that allows the
gateway to identify the process for retrieving the RSSI values following the
identifier, “NODE_RSSI_CID”.
Figure 4-12: Network Message Flow Sequence
53
Figure 4-12 demonstrates the process and the messages required to obtain the
inter-node RSSI throughout the SOSC network, containing the router nodes, end
devices, gateway, and the PC.
The topological structure and message sequence described previously
defines the communication network for obtaining the inter-node RSSI from every
SOSC node. Having described the messages and commands required to initiate
the SOSC devices to report the inter-node RSSI, the underlying flow of the
SOSC device firmware is summarized in the sequence illustrated in Figure 4-13
Main
(Awaits Sys Events)
afIncomingMSGPacket_t
= TRUE
?
No
OSAL Timer Expired
= TRUE
?
RSSI Request
CID = “PEG_ID_CID”
?
No No
Broadcast RSSI
Request
Yes
Yes
Forward RSSI value to
Gateway
Yes
Figure 4-13: Flow Chart of SOSC Router/End Devices
This main program is an infinite loop that constantly awaits the system event
interrupts passed from the OSAL layer. The occurrence of system events of
54
interest in this case are all indicated by the data buffer in
“afIncomingMSGPacket_t” structure. When this structure is found to contain the
system event, the main program continues to unpack and to check the system
event type, determining the process to be executed. To obtain the inter-node
RSSIs from the SOSC nodes, the two system event types of interest are the
timer event and the OTA message event. The timer event triggers the process to
broadcast the RSSI request and the OTA message event prompts the receiver
device to check the CID that initiates the forwarding of the RSSI value to the
gateway device.
The gateway device, being the listener that forwards the PC inter-node
RSSI, constantly awaits the OTA messages from the nodes and generates a
packet to be transmitted through UART. The UART packet follows the format
below
HEADER CID REPORT_NODE_ID RSSI RECEIVER_NODE_ID
2 Bytes 2 Bytes 2 Bytes 1 Byte 2 Bytes
Figure 4-14: UART Packet Format
The above packet is built and transmitted through UART on firmware level by the
gateway device. The flow of the gateway application is designed as show in
Figure 4-15
55
Main
(Awaits Sys Events)
afIncomingMSGPacket_t
= TRUE
?
No
Status Report
CID =
“NODE_REPORT_CID”
?
RSSI Request
CID = “NODE_RSSI_CID”
?
No No
Build Status Report
Packet and send
through UART
Yes
Yes
Build RSSI forwarding
packet and send through
UART
Yes
Figure 4-15: Flow Chart of Gateway Firmware
Similar to the firmware on routers and end devices, the gateway‟s application is
an infinite loop that awaits the OSAL system events, which in this case only the
OTA messages from the SOSC nodes. Upon the reception of the OTA
messages, the gateway identifies the two CIDs as either the status report
message or the inter-node RSSI message. The two CIDs define the two
processes that create the packet to be forwarded to the PC, based on the type of
message received.
The above concludes the design and description of the message structure
and communication topology in firmware level required for the SOSC WSN to
56
output the inter-node RSSI through the gateway device to the PC. The data
manipulation and the position computation from the localization algorithm on the
PC will be discussed in the next section.
4.3.2 PC Software Architecture:
Serial Port
Hardware
Serial Manager
Data ManagerLocalization
Algorithm
GUI
Software
Figure 4-16: PC Application Software Architecture
The PC application is written in C# and utilizes the Windows API to handle the
serial communication, such as USB and UART. This application‟s architecture
layout is shown in Figure 4-16, where the serial manager, using the predefined
function-calls from the Windows API, retrieves the RSSI and node information
data received in the data buffer. The RSSI and node information are to be
organized into the dissimilarity matrix table for the localization algorithm to
57
perform node position computation. The node positions ultimately will be
displayed with the GUI.
The data manager, upon each data reception event passed from the serial
manager, organizes the inter-node RSSI values into a matrix, indexed with the
RSSI‟s corresponding broadcasting node ID and the receiver node ID. The
algorithm for the data manager is shown as follows
Algorithm RSSI_Sorting(brctIDm, rcvIDm, RSSIm)
SOSC_nodelist[] =
Dissimilarmatrix =
i = 0
j = 0
for all m in M do
brctID = brctIDm
rcvID = rcvIDm
RSSI = RSSIm
if SOSC_nodelist[] = then
SOSC_nodelist[m] = Add_ID(brctID)
Obtain index of brctID in SOSC_nodelist[]
endif
if SOSC_nodelist[] ∩ rcvID = then
SOSC_nodelist[m+1] = Add_ID(rcvID)
Obtain index of rcvID in SOSC_nodelist[]
endif
i = Index_of(brctID in SOSC_nodelist[])
j = Index_of(rcvID in SOSC_nodelist[])
Dissimilarmatrix[i,j] = RSSIm
endfor
return Dissimilarmatrix
Figure 4-17: Data Manager Algorithm in the PC Application
The algorithm starts with initializing the array, SOSC_nodelist[], that keeps track of
the node devices joining the network; it then initializes a square matrix,
Dissimilarmatrix, to store the RSSI values. As mentioned earlier, with each
58
incoming data from the serial port, the algorithm checks if SOSC_nodelist[]
contains the broadcast node ID, brctID, and the receiver node ID, rcvID. New
nodes will be added to SOSC_nodelist[] if the corresponding IDs are not found in
the list. At the end, the RSSI value will be added to 2-dimensional Dissimilarmatrix
according to the indices of the brctID and rcvID in SOSC_nodelist[] respectively.
The process of storing RSSI values from the SOSC_nodelist[] indices into
Dissimilarmatrix is illustrated as follows
Figure 4-18: Data Manager Process for Constructing the Dissimilarity Matrix
The process illustrated in Figure 4-18 continues until all the inter-node RSSI
values from the SOSC WSN have been obtained to complete constructing
Dissimilarmatrix. This data manager not only allows the data to be stored in an
orderly fashion, but also allows the SOSC WSN to have the real-time knowledge
59
of the number of nodes currently in the network. Having had the dissimilarity
matrix for SOSC WSN with N nodes, the localization algorithm will determine the
node positions and store the position estimates in a 2xN matrix. The algorithm for
position computation is shown below
Algorithm MDS_Algo(Dissimilarmatrix[])
P = In – tileArray(
)
B = P x (-0.5(Dissimilarmatrix[])2) x P
Dissimilarmatrix[i,j] =
[E,V] = Find_EigenVect(B)
V = Sort(V)
for each n in N do
If Vn < 0
Vn = empty
else
Vn = Vn
endif
endfor
X =
return X
Figure 4-19: MDS Algorithm for Position Computation
60
Figure 4-20: Position Transformation Process
The algorithm shown in Figure 4-19 is the summary of Error! Reference source
ot found., (8), and (9). The details of matrix P and B are described in the
equations respectively. This algorithm is the portion of the localization algorithm
process, illustrated in Figure 4-20, and is responsible for transforming the inter-
node RSSI into the Cartesian-coordinate-based positions. The other portion of
the localization algorithm handles the mapping of those node positions and
acquires the correct layout in physical space. The design of the overall
localization algorithm will be discussed further in the next chapter.
To display the final mapping of the SOSC node distribution in space, a
GUI is designed to provide not only a visual understanding of the layout in space,
but also the real-time status of the SOSC WSN, such as the number of sensor
61
nodes within the network and their operating status. The architecture design for
the GUI is shown as follows
Main.cs
MainWindow
Position_DisplayNode_Status
Node(x,y) SignalNodeStatus
Signal
Data_Recorder
Node(x,y) Signal
Data
Set_Anchor
AnchorNodeID
Signal
Figure 4-21: SOSC GUI Architecture
The main program contains the localization algorithm and the I/O device handlers
built from the Windows API. The architecture shown in Figure 4-21 assumes the
incoming UART data has been received and processed in the background,
therefore the interconnection with the serial port is not shown. The “MainWindow”
is responsible for the visual display that handles the display of node position
layout, the data recording of the real-time localization computation results, and
the display of the real-time status of the SOSC nodes. Also, as part of the
localization computation, this GUI allows the user to choose the two anchor
nodes of interest to provide the correct visual layout of the SOSC nodes. The
GUI has the following look
62
Figure 4-22: The Actual GUI Implemented
Figure 4-22 illustrates the actual GUI used to display the result in localizing the
SOSC node positions. The GUI components highlighted with the red box are the
ones used for the purpose of localization and SOSC network related information
display.
4.4 Chapter Summary
This chapter describes the overall SOSC devices‟ network, hardware, firmware
architecture design, and the back-bone MDS algorithm design on a PC. As for
Node Status Display
Position Display
Set Anchors
Data Record
63
hardware architecture, two major components, the microcontroller unit and the
antenna are carefully chosen based on its functionality, such as the I/O capability
and antenna directionality, to reliably obtain distance data for localization
computation. The firmware architecture establishes the communication between
SOSC devices for the centralized monitor or computer, in this case, being the
PC, to obtain the inter-node distance information, enabling the PC application to
compute and to display the localization result. In the next chapter, we will discuss
the mapping algorithm to obtain the final localization result. The mapping
algorithm proposed in this thesis has a distributed nature. However, this thesis
concentrates on analyzing and demonstrating the localization performance and
not on the network performance. As a result, the algorithms are made centralized
on PC, consisting of the MDS based position computation algorithm, the RMU,
and the mapping algorithm.
64
Chapter 5 - Localization and Mapping Design
Previously, the software architecture of the MDS based position computation
algorithm was discussed. The proper RSSI-to-distance conversion model and
reconstruction of the SOSC node layout are yet to be described. Two of the most
important features of the SOSC WSN are the real-time ranging and the clustering
localization features. The real-time ranging is done through a range modeling unit
(RMU), which provides a simple ranging for position estimate. The clustering
localization allows the centralized localization algorithm to successfully and
thoroughly compute the location of every node within the network, regardless of
the scale and shape of the network layout. The detail of RMU and the clustering
mapping reconstruction features will be discussed in the following section
5.1 Range Modeling Unit (RMU) Distance Conversion
In real-world environments random electromagnetic interference and noise are
inevitable in radio transmission. These interferences introduce additional
degradation in signal strength, causing the RSSI values to degrade inversely
proportional to distance in a non-monotonic fashion. To illustrate this non-
monotonic inversely proportional RSSI to distance relationship, an experiment
was conducted by using two nodes to broadcast periodically to each other and to
obtain the RSSI of the broadcast packet at different distance away from each
other. The measurement starts at 2.5 meters separating the two nodes, and for
every 2.5 meter separation increments, 1000 RSSI samples are recorded over
time up to 40 meters.
65
The RSSI samples are averaged over the 1000 samples taken at each
distance and the final results are shown in Figure 5-1. It is obvious that, although
in general, the RSSI degrades in an inversely proportional fashion; at 27.5 and at
30 meters, the signals appear to be much weaker than that at further distances.
According to the literature [5] and [6], discussed earlier in Chapter 3 -, the
common cause for this weaker signal at 27.5 and 30 meters is due to fading or
various reflections from the surroundings.
Figure 5-1: Experimental RSSI to Distance Relationship
Regardless of the EM interference and fading that introduces irregularity in signal
strength degradation; there is still a clear inversely proportional relationship
between RSSI and distance according to Figure 5-1. Therefore, it is possible to
model this characteristic to allow the interpretation of distance from RSSI in an
ideal environment. However, in reality, without a proper radio channel model to
convert RSSI to distance, the accuracy of conversion will be easily influenced by
-90
-80
-70
-60
-50
-40
-30
-20
-10
0
0 5 10 15 20 25 30 35 40 45
RSS
I (d
Bm
)
Distance (m)
66
the ever-changing channel noise and interference. The loss in distance
conversion accuracy can result in a catastrophic localization estimation error.
Two common methods can be applied to conduct the RSSI-to-distance
conversion; the first one is the empirical method, where the system obtains
knowledge of the environment by collecting a sufficient amount of offline data,
such as the propagation property between any two nodes with known distance.
The empirical approach gathers the radio channel data for every link within the
network; the collection of data allows every link within the network to calculate a
probability density function (pdf) for channel properties, which allows the links to
determine the most probable propagation model to convert RSSI into the
corresponding distance. However, determining the radio propagation model for
every link requires the knowledge of all the inter-node distances, which defeats
the purpose of building a self-localizing and maintenance-free WSN.
The second method for RSSI-to-distance conversion is by utilizing the
theoretical radio propagation models. The advantage of using a theoretical model
is that it requires no system training and is easy to implement. However, one
theoretical model is unable to cope with all the random interference and
reflections in every environment. The theoretical models are only suitable in a
nearly ideal environment, where every node has perfect LOS communication in a
perfectly flat terrain without obstructions.
Apparently, the two conversion methods demonstrate different strengths in
different system design criteria. The former is suitable for a complex system with
good computation and storage capability for providing the most probable
67
conversion result from the large collection of off-line data. The second method is
simple, but it lacks the flexibility to adapt to random changes caused by various
interferences. As a result, this thesis proposes a hybrid solution that produces a
nearly-empirical and theoretical RSSI-to-distance conversion model in a run-time
environment. The nearly-empirical character of the self-calibrating distance
conversion module gathers the run-time inter-node RSSI and calculates the
propagation model in real-time. The propagation model is theoretical and is
based on (17), whose propagation constant, , and constant, , are calculated
upon each reception of each RSSI value. In the SOSC WSN, incorporating this
real-time RMU allows the system to interpret the RSSI values more reliably under
noisy conditions that cause the RSSI to appear weaker than predicted by the
theoretical model at a given distance. The following sections will describe the
RMU design and its performance compared to the empirical method.
5.2 SOSC RMU Architecture
The RMU‟s architecture is inspired by (17). The definitions of the two constants in
the equation define the communication scheme, facilitating the real-time radio
propagation modeling. The system architecture is illustrated as follows
68
Anchor 2
Anchor 1
Self-calibrating Unit
Sniffer
1 m
Fix
ed
Dis
tan
ce
Figure 5-2: SOSC Self-Calibrating Distance Conversion Unit Architecture
The RMU shown in Figure 5-2 consists of one “sniffer”, which is responsible for
retrieving the RSSI at 1 meter away from the neighboring anchor node. This
RSSI value represents the constant, , in (17). The other components included
by the RMU are the three anchor nodes. These three anchor nodes not only are
responsible for determining the correct mapping orientation and layout,
discussed later in the section, but also are responsible for determining the
propagation constant, . The three anchor nodes are fixed at some known
distance away from each other; the propagation constant is calculated from the
inter-anchor RSSI in the run-time environment. The reporting of RSSI values to
determine the propagation and the -constants are based on the asynchronous
69
periodic reporting scheme described previously in section 4.3. The flow for
calculating the real-time propagation model is shown below
PC App
Awaits
Packets
Sniffer RSSI
Packet?
Anchor RSSI
Packet?
Store A-Constant
Calculates n-
Constant
Awaits Anchor
RSSI Packet
Update
Propagation
Model
Yes
No
No
Yes
Figure 5-3: RMU Radio Propagation Model Forming Process
The process for determining the real-time radio propagation model, shown in
Figure 5-3, is a stand-alone process that operates in parallel with the process
that gathers the inter-node RSSI values from all SOSC nodes. This process of
RMU converts all the inter-node RSSIs into their corresponding distances in real-
70
time. It adapts to any interferences by updating the propagation model constantly
to ensure a correct distance conversion for each inter-node RSSI. However,
because noise distribution is uneven across space, it is difficult to characterize a
large network with just a single RMU. The performance of the RMU is
demonstrated in the following section.
5.3 Performance Comparison of the RMU
This section is to demonstrate how RMU is resistant to interference and to
successfully convert RSSI into the corresponding distance with minimal error.
This experiment is conducted with the set-up shown in Figure 5-2 in an indoor
environment, and the converted distances are recorded through the PC
application GUI. The set-up is illustrated below with a floor plan of the cubicle
Anchor 1
Anchor 2
Sniffer
PC App
Figure 5-4: Indoor Experimental Layout for RMU
71
This experiment is arranged in a way, shown in Figure 5-4 the two anchor nodes
are separated with the cubicle divider which can cause reflection and further
degrades the transmission strength at any given distance. The experiment is
conducted at 2 meters and at 7 meters separating the two anchors respectively.
During the experiment, the propagation constant and the -constant are
recorded. The average of the collection of propagation constants and the -
constants are used to define the empirical radio propagation model. Results of
the RMU conversion and the empirical conversion are compared and shown
below
Figure 5-5: Self-calibration vs. Empirical RSSI-Distance Conversion at 2m Apart
The objective for the two conversion methods is to provide the conversion result
as close to the physical separation, in this case, 2 meters, as possible. It is
1.5
1.7
1.9
2.1
2.3
2.5
2.7
2.9
3.1
3.3
3.5
0 10 20 30 40 50 60 70 80
Dis
tan
ce (
m)
# Samples
Real-time Calibrated Empirial Actual Distance
72
obvious in Figure 5-5 that the conversion by RMU can estimate the actual
distance better and more reliably than the empirical method. A detailed analysis
on the percentage error and variance of the conversion are shown in the
following table
Table 5-1: Average % Error and Variance of Distance Conversion with 2-meter separation
Average Error (%) Variance(m2)
RMU 6.9 0.03
Empirical Conversion 13.6 0.12
Table 5-1 shows that from numerous conversions performed at a fixed distance
of 2 meters, RMU almost doubles the accuracy, as compared to the empirical
conversion method. The variance of the conversion results obtained with RMU is
very small compared to that of the empirical conversion. These two indicators
demonstrate that RMU not only has a better RSSI-distance conversion accuracy,
but also has a much more consistent and a more repeatable conversion, due to
its lower variance. To further show RMU‟s superior capability in providing a
reliable RSSI-distance conversion, the same setup is used with separation of 7
meters. The results are shown in Figure 5-6
73
Figure 5-6: Self-calibration vs. Empirical RSSI-Distance Conversion at 7m Apart
Table 5-2: Average % Error and Variance of Distance Conversion with 7-meter separation
Average Error (%) Variance(m2)
RMU 8.8 0.58
Empirical Conversion 16.4 2.17
Table 5-2 shows the consistency in RMU‟s superior performance over the
empirical conversion. From conducting the above experiments in an indoor
environment with numerous obstructions, it is obvious that RMU is capable of
minimizing the influence of electromagnetic interference and noise to convert
from RSSI to distance with sufficient amount of accuracy. RMU, unlike the
empirical conversion, does not require off-line training and is able to adapt to the
environmental condition in real-time.
1.5
3.5
5.5
7.5
9.5
11.5
13.5
15.5
0 10 20 30 40 50 60 70 80
Dis
tan
ce (
m)
# Samples
Real-time Calibrated Empirical Actual Distance
74
5.4 SOSC Mapping Technique Design
As previously described, the procedure of a localization algorithm follows three
major steps; the first step is to gather distance information between each pair of
nodes within the network; secondly, the position computation algorithm
calculates the corresponding position in a 2D-Euclidean space; finally, the most
important step is to determine the actual layout for the node positions computed
in the former step. In [46], the author proposes a novel mapping technique by
partitioning WSN into quadrilaterals, formed by 4 nodes. The layout of the nodes
can be realized by finding the overlapping nodes between the quadrilaterals.
However, two of the most common problems in map reconstruction, also noted
by the author, are the translational and orientation ambiguity. As for orientation
ambiguity, it consists of rotational and flipping ambiguities. To reduce the number
of degrees of freedom, reference points or anchors are required to provide the
origin and direction of the map layout. In most cases, at least three anchors are
required to determine the layout. For simplicity, the following section will assume
the direction of the network layout and uses three anchor nodes to demonstrate
the procedure to overcome the above ambiguities.
5.4.1 SOSC Map Reconstruction
The main purpose of the map reconstruction is to align the MDS-estimated node
positions to the predefined anchor positions. This alignment procedure is an
iterative minimization problem which ultimately allows the anchor nodes to be
aligned to best fit their original configuration. As previously discussed, the ICP
algorithm facilitates this minimization of discrepancy between two sets of
75
coordinates. With the use of ICP (see section 2.3), the map reconstruction
algorithm is outlined in Figure 5-7
Figure 5-7: Map Reconstruction Procedure Flow
To demonstrate the reproducibility of the original mapping with the procedure
shown in Figure 5-7 and to justify the use of ICP algorithm for minimizing
mapping error, a simple scenario is created with the use of the following
dissimilarity matrix.
Obtain MDS-Estimated Node
Position
Rotate and Translate MDS-
Estimated Coordinates with
[R,T]
Output the Final Map
Calculate [R,T] = ICP(Estimated
Anchor, Actual Anchor)
76
Table 5-3: Dissimilarity Matrix of a Simple SOSC WSN
Anchor_1 Anchor_2 Node 1 Node 2 Node 3 Node 4
Anchor_1 0 0.5 0.4717 0.6021 0.3202 0.4924
Anchor_2 0.5 0 0.2693 0.4610 0.3905 0.5408
Node 1 0.4717 0.2693 0 0.2 0.2 0.2828
Node 2 0.6021 0.4610 0.2 0 0.2828 0.2
Node 3 0.3202 0.3905 0.2 0.2828 0 0.2
Node 4 0.4924 0.5408 0.2828 0.2 0.2 0
It is important to note that the inter-node distance information from Table 5-3 is
normalized for illustration purpose. The corresponding layout, shown with hollow
dots, of the above distance matrix is plotted with the MATLAB
Figure 5-8: Original Layout of the SOSC Network
This demonstration assumes that the translation and rotation vectors have been
calculated by ICP with respect to the two anchor nodes‟ original coordinates. The
Unknown Nodes Anchors
77
steps taken by the SOSC mapping algorithm to reconstruct the mapping are
illustrated below
Figure 5-9: Map Reconstruction Step One - Translation
The MDS algorithm, as described previously, provides the best matched solution
in terms of inter-node dissimilarity, but the corresponding estimated coordinates
very often do not reflect their actual layout. Therefore, to reconstruct the map, the
next step is to translate and rotate the MDS result to the designated anchor
positions
Translation
78
Figure 5-10: Reconstruction Step Two - Rotation
In Figure 5-10, it is easy to see that the current solution requires rotation to
completely position the two anchor nodes at their designated coordinates.
Figure 5-11: Reconstruction Step Three – Flipping
Reflection
Ground-Truth Positions
79
Figure 5-12: Reconstruction Step Four – Finalizing Mapping
Figure 5-11 demonstrates the duality of node layout. This duality occurs in a two-
anchor WSN or in situation where anchor nodes are collinear with one another.
Without prior knowledge of the network layout, it is difficult to determine whether
the estimated coordinates require flipping. To overcome potential flip ambiguity,
SOSC WSN utilizes three anchor nodes that are required to be not collinear.
This map reconstruction algorithm is independent of the number of nodes
within the network. The above example has demonstrated that this algorithm is
capable of reproducing the actual node layout from their inter-node distance
information. However, the above example is assumed to be in an ideal and fully
connected WSN without presence of EM interference that causes ranging error
induced localization error. Localization error can also be induced by partially
connectivity in a WSN, where a particular pair of nodes contains a broken link,
80
failing to produce distance information. This results in an empty entry in the
dissimilarity matrix and prohibits the classical MDS algorithm to fail. To address
the common range measurement error and connectivity issues that fail the
classical MDS algorithm, this thesis proposes a novel MDS-based localization
algorithm, called clustering MDS algorithm.
5.4.2 Clustering MDS Localization Algorithm
In ideal conditions, assuming perfect inter-node connectivity in the network
without EM interference, it is possible to localize the whole network with the
classical MDS algorithm from the inter-node distance information. However, in
reality, wireless nodes have limited signal coverage and cannot obtain the RSSI
from nodes outside of its coverage. This lack of connectivity between node-pairs
can cause nodes not to be localized. As discussed earlier, related works attempt
to use hop-count as their distance measuring metric; this technique can be
performed with multi-hop communication scheme and is less influenced by the
coverage. However, the disadvantage of hop-count based localization is its low
resolution and inaccuracy. As a result, for RSSI based localization, in order to
cope with the nodes‟ inability to obtain distance information from out-of-range
nodes, this thesis introduces a novel localization method based on the concept of
clustering.
The design of clustering MDS localization algorithm is inspired by the previous
work done by Stefano Basagni [47] and Rajesh Krishnan [48]. Their works apply
the concept of clustering to group the nodes in an ad-hoc network to create a
hierarchical network organization to reduce the amount of data required to
81
maintain routing and control information in a mobile environment. Each cluster is
governed by a cluster-head, defined by the node having the largest weight; the
nodes within a one-hop distance away from the cluster-head will become
members of the cluster. Based on the concept of clustering, in localization
applications, many ideas from [49], [50], and [51] also apply local information to
construct the global map of WSNs, whose nodes do not have perfect connectivity
with every other node in the network. The classical MDS algorithm can be
applied in the cluster by cluster fashion, so it does not require knowledge of the
inter-node distance information of the global map. As a result, even without
connectivity of certain node-pairs, the clustering MDS localization is still able to
localize the particular node from its other neighboring nodes with good
connectivity. The algorithm for the clustering MDS localization is
Algorithm Clustering_MDS(Anchor_Pos{xi,yi} i=1,2,3 , Dissimilarmatrix[])
Number_of_Network_Nodes = length_of(Dissimilarmatrix[])
node_to_clutDist =
Clust_Index = {1,2,3}
If (N != 0)
for each n in N do
for each m in length_of(Clust_Index[]) do
node_to_clutDist[1,n] += Dissimilarmatrix[n, Clust_Index[1,m]]
endfor
endfor
Closest_Node = min(node_to_clutDist)
Closest_ThreeAnchors =
Find_Three_MinAnchor(Dissimilarmatrix[n, Clust_Index])
Closest_ThreeAnchors = Add(Closest_Node)
New_Dissimilarmatrix =
for j = 1 to 4 do
for k = 1 to 4 do
if j k then
82
New_Dissimilarmatrix[j,k] =
Dissimilarmatrix[Closest_ThreeAnchors[1,j], Closest_ThreeAnchors[1,k]]
endif
endfor
endfor
Est_Node_Positions = MDS_Algo(New_Dissimilarmatrix)
[R,T] = ICP(Anchor_Pos, Est_Node_Positions)
Map Noden with R and T
Anchor_Pos = Add(Noden)
N = N-1;
Clust_Index = Add(Index_of(Noden in Dissimilarmatrix))
else
return Anchor_Pos
endif
Figure 5-13: Clustering MDS Algorithm
At the start, the WSN contains N un-localized nodes and three anchor nodes with
known physical locations. The three anchor nodes form the initial cluster for the
network. During run-time, the algorithm, for each un-localized node, adds all the
inter-node distances to its cluster nodes and determines the un-localized node
that is closest to the current cluster. This closest node will be localized by its
three closest cluster nodes with the classical MDS algorithm. Once the node is
localized, ICP algorithm will be used to calculate the corresponding rotation and
translation required to map the estimated location of the un-localized node with
respect to the closest three cluster nodes. After conducting rotation and
translation, the node will be added to the existing cluster and used to localize the
rest of the un-localized nodes.
According to the RSSI range data collected with a pair of nodes, the result
in Figure 5-1 has demonstrated a non-monotonic degradation in RSSI in regions
between 25 and 40 meters. The propagation model in (17) shows that as
83
distance increases, RSSI value‟s rate of attenuation decreases. When
interpreting distance with weak RSSI values can introduce a large and distance
estimation error. Figure 5-14 is the illustration of ranging error induced by using
weak RSSI signals as sources. It is clear to see that the same 30 dBm noise that
degrades a stronger RSSI signal and a weaker RSSI signal results in completely
different ranging errors. In this case, the stronger RSSI introduced only 3 meters
of error, while the weaker RSSI introduces 30 meters of discrepancy. As a result,
to reduce the error induced by EM interference or fading, the clustering MDS
algorithm described earlier, specifically localizes the node that contains the
strongest RSSIs to the cluster nodes. In other words, the localization algorithm
proposed by this thesis performs localization based on the fact that weaker signal
is more prone to error in distance estimate. For nodes that are closer together,
the RSSI values tend to be stronger and their path loss characteristic is much
more monotonic, and therefore providing a more accurate distance information
and mitigating the overall localization error.
84
Figure 5-14: RSSI Error vs. Ranging Error
5.5 Chapter Summary
This chapter describes the design of the RSSI to distance conversion unit, RMU,
the mapping methodology, and finally, the design of the clustering MDS
localization algorithm. Since, the SOSC WSN is designed for low-cost and low-
power applications; it does not require complex antenna and timing systems
required for the AOA and TDOA ranging techniques. Instead, it adapts a RMU,
which utilizes a generic propagation model to characterize the real-time distance
conversion model in a real-time environment. Due to random EM interferences
and ever changing climate conditions, the need for RMU is shown to be
essential, because an empirical conversion model cannot dynamically adjust
30 dBm
3 m 30 m
30 dBm
85
parameters in the propagation model to account for the changes in the
communication channel. Having had the distance information, this thesis
demonstrates that it is possible to reconstruct the physical map by translation
and rotation, with only three anchor nodes. Although, RMU is used to provide a
real-time RSSI to distance interpretation, it is not capable of detecting the faulty
RSSI values, induced by fading or interferences. As a result, clustering MDS
algorithm is designed to discard the weaker RSSIs, which are prone to ranging
error, from localization. Clustering MDS algorithm only considers the stronger
and more reliable RSSIs to estimate node locations and so is able to mitigate
localization error. In Chapter 6, simulations and experimental results will be
shown to proof the validity of the above claim.
86
Chapter 6 - Results and Discussion
As previously discussed, due to limitation in radio signal coverage, nodes in a
large scale WSN are not able to obtain distance information from all other nodes
in the network. Also, due to EM interference and fading, the inter-node distance
information obtained from network does not always reflect to the physical
distance. The missing connectivity and faulty distance information can lead the
classical MDS algorithm to provide poor position estimation, resulting in
erroneous global map reconstruction. RSSI ranging is not known for its reliability
and accuracy, but many literatures, such as [52], have attempted to perform
localization with only radios so the overall system can be made inexpensive and
energy efficient. Essentially, the accuracy of the localization is limited to the
accuracy of the range measurement. Authors in [53] enhance this inaccuracy by
performing an extensive data collection to build a statistical range model for
various types of environments. All the literatures this thesis has discussed earlier
attempts to achieve a good localization result by incorporating extensive number
of anchor nodes, refined ranging techniques, or localization computation that rely
greatly on the knowledge of the network environment. As for SOSC WSN, the
system requires only three anchor nodes for mapping reference; a simple RMU
that characterizes the real-time RSSI-to-distance model; a simple clustering MDS
algorithm that can localize partially connected WSN with much less computation
complexity than the classical MDS localization. In this chapter, simulations and
real-world scenarios on localization and computation complexity will be
demonstrated to show its superiority over the classical method.
87
6.1 Computation Complexity
The classical MDS algorithm is computationally expensive in nature. It is mainly
due to Eigen-decomposition, whose computation complexity is . To localize
a large scale network, using the classical MDS algorithm can be computationally
expensive. The clustering MDS algorithm, on the other hand, performs the
classical MDS algorithm in a cluster of 4 nodes for clusters, excluding the 3
anchor nodes. The resulting computational complexity without the mapping
reconstruction algorithm is , where is the constant time required to
perform the classical MDS over 4 nodes. The map reconstruction algorithm, in
this case being the ICP algorithm, whose computation complexity is ,
performs linear iteration over 4 nodes and repeats for clusters. This
process iterates through each individual un-localized node across the whole
network. As a result, the overall clustering MDS mapping algorithm has
computational complexity of , where is the constant time
combining the computation time of 4-node MDS and 4-node ICP algorithms,
which is simply .
The classical MDS algorithm exhibits a computation complexity of ,
while the clustering MDS algorithm only produces a complexity of . The
clustering MDS algorithm, due to its linear computation complexity, is much more
scalable for large network than the classical MDS algorithm. By examining the
computation complexity of the two algorithms, the clustering approach has
already established a great advantage over the classical method.
88
6.2 Localization Performance
In this section, simulations are conducted to compute localization performance of
classical MDS and clustering MDS in both ideal and noisy environments. The
simulation of an ideal environment assumes perfect range measurement and
perfect node connectivity within the network. All localizations are performed with
only two anchor nodes, placed along y-axis for mapping reference. The
localization results are distributed in a 1x1 unit square space for illustration
purposes. The simulation results for classical MDS and clustering MDS are
shown in Figure 6-1
Figure 6-1: Classical MDS Map Reconstruction
The nodes marked with „○‟ are the actual layout and ones marked with „x‟ are the
estimations. This simulation is done by randomly generating the fifty node
positions at start-up. The inter-node distances are calculated directly from the
89
generated coordinates, because this situation is assumed to be ideal, so the
distance information should reflect its actual physical layout. The localization
results are calculated with the MDS algorithm from the inter-node distance
matrix. The objective of this simulation is to show the MDS algorithm‟s capability
to reconstruct the original layout from just the distance information. It is easy to
see that the estimation match with the original layout perfectly. Another
simulation, shown below in Figure 6-2, uses clustering MDS localization
algorithm.
Figure 6-2: Clustering MDS Algorithm Map Reconstruction
From Figure 6-2, it is obvious to see that, with the clustering MDS algorithm, the
localization result matches with the original layout. However, just by visualizing, it
is difficult to evaluate the performance of the two algorithms. Therefore, a
quantitative measure is required for evaluating the two algorithms. In this case,
90
mean error of position estimation is calculated for individual nodes within the
network. The estimation error is defined by the following equation
(18)
Where is the number of nodes within the network, and the Euclidean distance is
measured with respect to the origin. This error essentially evaluates the
“closeness” between the original point and the estimated point as a fraction of the
original point‟s distance away from the origin. Utilizing this formula, the estimation
accuracy of both the classical MDS and clustering MDS algorithms under perfect
environment conditions are shown in Table 6-1
Table 6-1: Position Estimation Error Comparison
Algorithm Type Estimation Accuracy (%)
Classical MDS Algorithm
Clustering MDS Algorithm
It is obvious that under perfect environmental and network conditions, both
algorithms are able to reconstruct the original map layout with minimal error.
However, the clustering MDS algorithm in this case exhibits a slightly worse
accuracy then the classical method. This deviation in estimation accuracy is due
to various round-off and precision errors during the rotational and translational
map transformation processes. A larger propagation error can exist for a larger
WSN. However, the round-off and precision error compared to the distance
estimation error in real-world scenario is negligible. The following simulations
demonstrate the estimation error of the two localization algorithms under more
91
realistic conditions, where the inter-node distance information is influenced by
EM interference and induces error.
The experimental data in Figure 5-1 shows that as distance between
nodes increases, the inter-node RSSI weakens, but the RSSI‟s rate of decrease
is inversely proportional, resulting in poor resolution for interpreting far inter-node
distances. Radio signals are prone to noise and fading in nature, and therefore, a
slight variation in weak RSSI can introduce a large variation in distance
conversion. To demonstrate the effect of signal noise on localization, a simple
model on noisy distance information is defined as
(19)
where , , and is total number of nodes within the network. This
simple model assumes that the noise in RSSI caused by interference or fading
scales linearly with distance. The further an inter-node distance is, the larger the
distance measurement error. In reality, this error relationship can be exponential
according to the observation from Figure 5-1. However, to simplify calculations
and for illustration purposes, it is sufficient to assume a linear distance estimation
error model.
To demonstrate the above claim, two simulations are presented with ten
unknown and randomly distributed nodes; maximum range error of 13% and 23%
are introduced respectively to the nodes. The localization results calculated from
classical MDS and clustering MDS algorithms are shown below
92
Figure 6-3: Classical vs. Clustering MDS Localization with 13% Range Error
The mean estimation error is the average of all the estimation error of every node
in the network. It is not surprising to see that the classical MDS algorithm
produces a larger positioning error than the clustering MDS algorithm. The
classical MDS algorithm calculates the best-fit map solution based on all the
inter-node distance information within the network. In the above simulation,
where the range error is bounded by the maximum error of 13%, the classical
MDS algorithm has to account for every error, and therefore its solution will
approach an error of 13%. On the other hand, the clustering MDS algorithm
always progressively localizes the closest node, whose distance information is
less influenced by variation in RSSI, so its localization result is not dependant on
the maximum error. In other words, because of the data-selective nature of
clustering MDS, poor distance estimates are discarded from the localization
calculation, and the overall positioning error can be further reduced.
93
The above observation is further verified with the following simulation with
a larger range measurement error of 23%. The results are shown in Figure 6-4
Figure 6-4: Classical vs. Clustering MDS Localization with 23% Range Error
It is easy to see that as the range error increases, the overall mean positioning
error increases and approaches the maximum range error within the network.
The same distance information is applied to the clustering MDS algorithm and its
localization result is shown in the right side of Figure 6-4. The results for both
algorithms demonstrates consistency in performance, where the classical MDS‟s
positioning accuracy is always limited by the largest error in distance estimate;
the clustering MDS‟s positioning accuracy tends to minimize the range error by
always localizing nodes with less error relative to other nodes in the network.
94
6.3 Experimental Localization
6.3.1 Indoor Localization
In the previous section, a noisy distance model is defined in (19), based on
observation from RSSI range measurement. The localization algorithms calculate
the node positions from noisy distance information. From the localization results,
it is clear to see that without the clustering MDS approach; the localization‟s
accuracy is limited by the largest range measurement error of the dissimilarity.
To support this claim further, a simple real-world scenario with 4 nodes is set up
indoors with a node layout depicted in Figure 6-5
GatewaySniffer
Node 3
Node 4
Node 7
Node 8
5 m
5 m
1 m
Figure 6-5: Indoor Localization Setup
95
Figure 6-6: Localization Results from Clustering MDS and Classical MDS
Table 6-2: Positioning Error of Localization
Node Classical MDS Estimated Position
Classical MDS Estimate Error
(%)
Clustering MDS Estimated Positions
Clustering MDS Estimate
Error (%)
Anchor Node #8 (0,0) 0 (0,0) 0
Anchor Node #4 (0, 4.97) 0.58 (0, 5) 0
Node #3 (3.28, 1.25) 42.5 (3.67, -0.32) 27.3
Node #7 (2.48, 2.54) 70.4 (2.91, 2.39) 66.9
The setup consists of 4 nodes, including 2 anchor nodes with known positions;
the nodes are placed at four different corners of the couches, measuring 5 meter
in length. Moreover and RMU is also present as part of the network to perform
real-time RSSI to distance conversion. The result is shown in Figure 6-6 and the
corresponding estimation error is listed in Table 6-2. Using the same set of
dissimilarity collected, localization is further conducted and verified, using
96
different anchor nodes as the mapping reference. The result is shown in Figure
6-7 along with the position error in Table 6-3
Figure 6-7: Localization Results from Clustering MDS and Classical MDS using Different
Anchors
Table 6-3: Positioning Error for Different Anchor Nodes
Node Classical MDS Estimated Position
Classical MDS Estimate Error
(%)
Clustering MDS Estimated Positions
Clustering MDS Estimate
Error (%)
Anchor Node #3 (5,0) 0 (5,0) 0
Anchor Node #7 (5, 5) 0 (5, 5) 0
Node #4 (1.46, 3.48) 42.1 (1.4, 3.47) 41.6
Node #8 (7.68, 2.46) 161.2 (2.24, 2.44) 66.3
From the above two scenarios, it can be seen that the results calculated with
classical MDS method are very different under different network configuration.
First, in Figure 6-6, applying the classical MDS algorithm with respect to anchors
97
Node 8 and 4, introduces a maximum positioning error of 70%. Because Node 7
has the worst estimation error of all, the overall localization is compensated for its
error. On the other hand, the clustering MDS, although, is unable to improve the
position estimate of Node 7 further, Node 3‟s location is calculated independently
from the inter-node distances with Node 7 and 8, ignoring the more erroneous
distance information of Node 4. The final positioning error for Node 3 is much
less when localized with the clustering MDS algorithm than with the classical
MDS algorithm. The layout in Figure 6-7 also demonstrates the superiority of
clustering MDS localization and is consistent with the simulations shown
previously. However, the clustering MDS localization cannot achieve high
accuracy, because under random EM interference and reflections, it is extremely
difficult to obtain stable RSSI values for reliable characterization with RMU.
From previous localization in indoor environment, it is obvious that the
disadvantage of classical MDS localization is that it only provides one solution
based on the dissimilarity provided. Therefore, when the dissimilarity contains
range measurement error, the localization accuracy is always compromised by
the largest error within the network. On the other hand, with the clustering MDS
algorithm, nodes having larger error are often rejected from the position
computation, and therefore resulting in a position estimate with better accuracy.
In the next section, an outdoor distribution of SOSC network is deployed to
further demonstrate clustering MDS algorithm‟s ability to reduce the localization
error.
98
6.3.2 Outdoor Localization
Figure 6-8: Node Layout on the Field
In the previous section, due to limited indoor space for node distribution, only four
nodes were placed and tested. This section, the SOSC system and its algorithm
is further tested in a larger scale outdoor environment with total of seven nodes.
There are a number of objectives to be achieved and verified; the objectives are
To demonstrate the affect of range measurement error on localization
accuracy
To demonstrate the affect of network density on localization accuracy
To compare the localization performance between classical MDS and
clustering MDS algorithm
Figure 6-8 show a grass field at SFU, where the SOSC nodes were placed for
the experiments. To achieve the objectives listed above, three layouts of different
99
network density, high, medium, and low density were used. In each density
layout, inter-node distances are measured with a tape measure. Table 6-4 is an
example of the inter-node distances from a high density node layout. In order to
compare the localization performance at the end of each test, ground truth node
positions have to be defined as reference. This ground truth is obtained from
performing classical MDS algorithm on the tape measured dissimilarities.
Table 6-4: Tape Measured Dissimilarities of SOSC Nodes
Tape Measure Dissimilarities (m)
Node ID 2 3 6 7 9 10 11
2 0 13.7 24 10.9 19.7 12.7 20.6
3 13.7 0 15.7 13.2 18.5 23.6 26.5
6 24 15.7 0 15 10.9 26.9 23.1
7 10.9 13.2 15 0 8.8 12.4 13.4
9 19.7 18.5 10.9 8.8 0 17.7 12.2
10 12.7 23.6 26.9 12.4 17.7 0 10.7
11 20.6 26.5 23.1 13.4 12.2 10.7 0
Given the transmission range of the node radio, the maximum inter-node
distance was kept within 60m. The area coverage of the high, medium, and low
density networks were designed to be 20mX25m, 50mX25m, and 60mX30m,
respectively. To compute the ground truth coordinates, dissimilarities in Table 6-4
were computed with MDS algorithm. The network layouts are shown in Figure
6-9, Figure 6-10, and Figure 6-11
100
Figure 6-9: High Density Ground Truth
Layout
Figure 6-10: Medium Density Ground Truth
Layout
Figure 6-11: Low Density Ground Truth Layout
Once positioned, every node location is measured with a commercial handheld
GPS device for comparison purpose. A GPS reports coordinates in terms of
degrees of latitude and longitude. To represent the GPS coordinates in Cartesian
coordinates, a standard WSG84 conversion, [54], is used. The resulting
101
Cartesian coordinates of the seven nodes, compared with the ground truth layout
are shown in Figure 6-12, Figure 6-13, and Figure 6-14
Figure 6-12: High Density GPS Localization
Comparison
Figure 6-13: Medium Density GPS
Localization Comparison
Figure 6-14: Low Density GPS Localization Comparison
It is important to note that the commercial GPS used for this test was not a
differential GPS, and therefore it did not have the capability to correct its
102
measurement error due to the change in atmospheric conditions at the moment
of measurement. This is one of the major factors causing the large discrepancy
for node 6, 9, and 11 in the above three figures. Besides those localization
errors, the overall GPS position estimation accuracy is listed in the table below
Table 6-5: Average GPS Localization Error
GPS Localization Error
Density High Medium Low
Average Distance Error (m) 6.08 6.1 12.63
It is obvious to see that even with state-of-art commercial GPS devices, it is still
difficult to localize with high accuracy.
Having demonstrated the GPS localization performance, the experiment
continued with RSSI-based localization with both classical and clustering MDS
algorithms. The test utilized the RMU (see section 5.1) to convert real-time inter-
node RSSI into corresponding inter-node distance. Each network layout is
operated until a total of 1000 inter-node RSSI and distance have been received
and converted. The RSSI-distance model is ,
where is 2.5 and is 32dBm. At each density, the RSSI measured at each
inter-node link is compared with its corresponding ground truth distance. The
resulting plots for each layout density are shown below
103
Figure 6-15: RSSI vs. Distance (High
Density)
Figure 6-16: RSSI vs. Distance (Medium
Density)
Figure 6-17: RSSI vs. Distance (Low Density)
From the above three figures, it is clear to see that RSSI values do not follow
their theoretical values at each distance. The RSSI values in the high density
case, due to noise and RF interference, spans 10‟s of dBm at the distance
marked by the dotted line. As the networks became less dense, the inter-node
RSSIs became more distinct at each distance and started to merge toward the
RSSI-distance model more closely.
104
Using the above dissimilarities, the localization results computed by
classical MDS and clustering MDS at each layout density are shown below
Figure 6-18: High Density Classical MDS
Result
Figure 6-19: Medium Density Classical MDS
Result
Figure 6-20: Low Density Classical MDS Result
Anchors
105
Placing the nodes and obtaining ground-truth data via tape measure is very time
consuming. Therefore every possible anchor configuration for each density was
evaluated. This is done by choosing three anchor nodes out of the seven nodes,
and it gives us possible network configurations for each density, making
a total of 3x35 = 105 test configurations. The figures shown above are the best
localization result from the 35 separate tests of each density. The table below is
a summary of the average localization accuracy in each network density.
Table 6-6: Classical MDS Localization Accuracy
Classical MDS Localization Error
Density High Density Med Density Low Density
Average Error (m) 15.51 28.12 21.43
Min Error (m) 8.29 16.56 13.56
Max Error (m) 20.02 37.37 31.67
STD (m) 2.64 5.57 4.2
Notice that the localization error is larger in the medium density layout than in the
low density layout. This is because the non-monotonic degrading behavior of
RSSI with distance in the medium density, Figure 6-16, did not follow the
theoretical model as closely as that in the low density scenario in Figure 6-17.
This example demonstrates the fact that although, node density plays an
important role in localization accuracy, localization accuracy is greatly dependent
on range measurement accuracy.
Having had explored the localization performance of the classical MDS
algorithm, the localization performance of the clustering MDS algorithm is tested
and analyzed, following the same procedures described previously. The results
are shown in Figure 6-21, Figure 6-22, and Figure 6-23
106
Figure 6-21: High Density Clustering MDS
Result
Figure 6-22: Med Density Clustering MDS
Result
Figure 6-23: Low Density Clustering MDS Result
107
Table 6-7: Clustering MDS Localization Accuracy
Clustering MDS Localization Error
Density High Density Med Density Low Density
Average Error (m) 13.19 21.68 23.71
Min Error (m) 6.17 10.66 15.91
Max Error (m) 21.56 34.85 36.56
STD (m) 3.65 5.12 5.02
From the above three network densities, a total of 105 localization tests have
been generated. To summarize the overall localization performance with respect
to the node density, the average of the localization results under the same level
of node density is calculate for every test case. The final localization error versus
node density comparison is shown below
Figure 6-24: Localization Accuracy vs. Node Density
In general, Figure 6-24 shows that the localization accuracy is directly
proportional to the network density and that clustering MDS exhibits an
108
improvement on the accuracy compared to the classical MDS. However, the
clustering MDS algorithm, in situations where certain node pair‟s RSSIs are
constructively interfered and appear to be closer to each other, the estimated
location of that particular node pair can result in a large discrepancy, like node 6
in Figure 6-23. The nature of clustering MDS allows improvements on localization
result in fading situations, like in the high and med-density scenario; it does not
perform better than the classical MDS algorithm when RSSIs are constructively
interfered in a lower density network, as illustrated in the lower density portion in
Figure 6-24.
It is a well known fact that radio wave‟s signal strength is easily degraded
by EM interference. As observed from the experiments, RSSIs can be either
destructively or constructively interfered to produce values that do not
correspond to their physical distance. This behavior has been shown in Figure
6-15, Figure 6-16, and Figure 6-17 and been shown to result in range
measurement errors. To investigate the influence of range measurement errors
on localization performance, an average range error is calculated from difference
between the RSSI converted dissimilarity with the ground-truth dissimilarity. The
localization error is then plotted against its corresponding range measurement
error. The results from both the classical and clustering MDS are shown below
109
Figure 6-25: Localization Error vs. Range
Error (Classical MDS)
Figure 6-26: Localization Error vs. Range
Error (Clustering MDS)
Being largely dependent on range measurement to perform localization, it
is clear to see that in Figure 6-25 and Figure 6-26, the localization error
distribution, represented by the error bars, of the two algorithms increases
proportional with range error. However, due to the clustering MDS‟s selective
nature on choosing distance information to localize nodes, the overall localization
accuracy is less susceptible to erroneous distance information than the classical
MDS algorithm. From the previous observations, the clustering MDS algorithm
has an advantage over the classical MDS algorithm in medium and high density
networks consisting of erroneous distance information. However, to justify one
algorithm‟s superiority in localization performance over the other; it is important to
further show the uniqueness by investigating the statistical distribution of the
localization accuracy from the two algorithms.
110
Figure 6-27: Classical MDS Localization Error Distribution
Figure 6-28: Clustering MDS Localization Error Distribution
From the 105 localization tests, the localization error of each algorithm is
summarized in the two histograms, Figure 6-27 and Figure 6-28 respectively. To
determine the similarity between the two statistical distributions before
concluding on the uniqueness of clustering MDS algorithm, a two-sample t-test is
conducted for each network density and for the overall case. The significance of
each case is shown in the table below
111
Table 6-8: Two-sample T-Test of Statistical Significance
T-Test
Significance
High Density 3.30E-03
Medium Density 3.61E-06
Low Density 0.044
Overall 1.83E-02
By definition, statistical similarity between two distributions is indicated by the null
hypothesis, which states that the two distributions are of the same statistical
mean, can be rejected if the significance is below 5%. From Table 6-8, it is clear
to see that in every scenario, the clustering MDS algorithm is indeed a different
algorithm from the classical MDS.
6.3.2.1 Network Connectivity vs. Localization Error
In previous test cases, perfect connectivity of inter-node communication was
assumed. In fact, in medium and low density network cases, there existed pairs
of nodes that were not able to communicate with each other and therefore no
RSSI values for these pairs were recorded. When one node fails to receive RSSI
from the transmitting node, the dissimilarity matrix can result in an empty entry
for that particular node pair, and therefore leaving the dissimilarity matrix
unsymmetrical. Classical MDS algorithm cannot solve an unsymmetrical matrix
by default. A proper scheme to handle the disconnected graph in a real network
is expected, otherwise it is impossible to localize using the classical MDS
algorithm.
112
Unlike classical MDS, which requires each individual inter-node distance to
performing localization, the clustering MDS algorithm only requires the distance
information of the closest neighbor to the anchors that are performing localization
at the time instant. As a result, the algorithm ignores the disconnected link and
utilizes other available links to localize. In this section, an example of
disconnected link from the medium density case is re-created to demonstrate the
localization performance of the clustering MDS algorithm over the classical MDS
algorithm.
The following table is the dissimilarity collected from nodes in the medium
density layout. The high-lighted node pair had a poor reception and was unable
to produce a symmetrical matrix.
Table 6-9: Unsymmetrical Dissimilarity
Medium Density Dissimilarities (m)
Node ID 2 3 6 7 9 10 11
2 0 19.46 12.24 17.59 31.79 16.87 55.51
3 19.46 0 23.66 55.36 15.92 10.87 22.01
6 12.24 23.66 0 47.52 50.33 33.84 30.76
7 17.59 55.36 47.52 0 42.75 40.23 0
9 31.79 15.92 50.33 42.75 0 21.80 8.60
10 16.87 10.87 33.84 40.23 21.80 0 15.57
11 55.51 22.01 30.76 91.63 8.60 15.57 0
In this case, a simple algorithm can be implemented to regenerate a symmetrical
matrix by determining one of the missing entries and filling it with the distance
information received by the other node. However, in a more extreme case, it is
possible that this node pair cannot receive distance information in either
113
direction. To simulate this scenario, the high-lighted cells in Table 6-9 are both
filled with 0 to indicate a complete broken link. This dissimilarity is then localized
by both the classical and clustering MDS algorithms. The same test to produce a
comparison between average node densities and localization error as shown in
Figure 6-24. The test was conducted for both the fully connected case and the
disconnected case; the results are shown as follows
Figure 6-29: Localization Accuracy in Disconnected Network
114
Figure 6-30: Localization Accuracy in Fully Connected Network
By removing one entry from the dissimilarity matrix, the average localization error
in the disconnected scenario in Figure 6-29, has worsened by almost 15 meters
on average. On the other hand, clustering MDS demonstrated only a slight
degradation in performance. In this particular test, the WSN only consists of one
broken link; in a larger scale and less dense network, the number of
disconnected links can increase, causing the classical MDS algorithm to fully
localize the network. By conducting this test, it can be verified that the clustering
MDS is capable of withstanding the influence of missing inter-node distance
information, sustaining the same level of accuracy as performing localization on a
fully connected network.
115
6.4 Chapter Summary
In this chapter, based on computation complexity, simulation, and experimental
results, we have outlined the advantages of localizing with the novel clustering
MDS algorithm over the classical MDS algorithm. Conducting the tests with three
different network density layouts, we were able to produce 105 individual network
configurations. The 105 tests consist of different levels of node density and
ranging errors. We summarized the results from these test results and analyze
localization performance between the classical MDS approach and the clustering
MDS approach against node density, ranging error, and connectivity of the
network. The clustering MDS localization is shown to be capable of producing
comparable and more resilient localization accuracy under noisy condition and
poor network connectivity. Moreover, clustering MDS also demonstrates a much
lower computation complexity, which ultimately makes this algorithm truly
suitable for low-power and low-cost WSN localization applications.
116
Chapter 7 - Conclusion
In this thesis, we acknowledged the major difficulties in assuring performance in
localization systems to be the hardship in defining the ground truth reference and
the range measurement error. We have demonstrated the former by comparing
the localization results between the GPS measured node positions and the MDS
calculated node positions based on tape-measured distances. It is obvious that
even with the commercial GPS device; the localization accuracy can still
introduce discrepancies. The second difficulty is the most common in all range-
based localization systems. Many previous works, discussed earlier, cope with
the ranging error by off-line training or by developing complex models to
minimize error in range measurement. It is also common to equip sensors with
additional ranging devices to reduce the biasing effect, induced by using only one
ranging device. Most of the solutions from related works are costly and complex
to implement. As a result, in this thesis, we developed a low-cost and low-power
SOSC system for ZigBee WSN to address the above problems and to also
deliver comparable accuracy and much lower computation complexity than the
classical MDS localization algorithm.
Having said ranging error is a major cause to degradation of localization
accuracy, we introduced a real-time RMU to facilitate the distance conversion
process. As opposed to the empirical approach, this RMU is adaptive to the ever-
changing channel parameters in real-time and does not require training for the
system. From experiments conducted in a small indoor environment, we were
able to show that the RMU estimates the true distance better than the empirical
117
approach. However, because RMU converts RSSI to distance only based on
three anchor nodes and one sniffer in an attempt to characterize the whole
network. In reality, due to uneven noise distribution in space, it is challenging to
generalize conversion models. However, to provide a more accurate conversion
model to represent a large network, more than one RMU can be deployed to
characterize individual region in the network.
Although, using only RSSI ranging for localization introduces limitation on
producing highly accurate node position estimate, we proposed the clustering
MDS along with the ICP algorithm to minimize mapping error. We also
demonstrated the clustering MDS localization algorithm‟s ability to cope with the
influence of ranging error and imperfect network connectivity; the clustering MDS
is able to produce comparable localization results with much less computation
complexity. Being linear to the scale of the network, the computation complexity
of the clustering MDS allows the SOSC WSN to be cost-effective and attractive
to large network applications.
7.1 Future Work
Future works to further improve this SOSC WSN include implementing the
clustering MDS localization algorithm in firmware (in a distributed way) on low-
cost and low-power microcontrollers; second, a more refined ranging technique
and accurate distance conversion modeling are required to enhance the
localization accuracy.
First, the clustering MDS algorithm is a distributed algorithm by nature;
however, the focus of this thesis is on analyzing its improvements on localization
118
accuracy and computation complexity over the classical MDS localization.
Therefore, the algorithm is implemented in a centralized manner and the benefit
of reduced network overhead, delivered by using distributed algorithms in WSN
is not discussed. For future work, this localization algorithm can be modified and
implemented on the CC2530 SOC with an improved arithmetic and further
reduced complexity to maintain low-power operation.
Second, as observed previously, RSSI ranging introduces large error,
degrading the overall localization performance. To improve the current state,
more distance information encapsulated by RF signals can be extracted to
ensure the reliability of the range measurement without introducing additional
ranging devices. For example, in most ad-hoc networks, hop-count is a popular
feature and can be embedded in RF signals in inter-node communications. Hop
counts are essentially a representation of communication cost; communication
cost can be influenced by interference or natural signal degradation due to
distance, but it is less sensitive than RSSI. Hop-count can be utilized in
conjunction with RSSI for ranging purposes. By having this additional
information, correctness of the range data can be determined more reliably, and
the clustering MDS localization algorithm can be less biased by always localizing
nodes with strongest RSSI.
119
References
[1] N. Bulusu, J. Heidemann, and D. Estrin, "GPS-less low-cost outdoor localization for very small devices," Personal Communications, IEEE, vol. 7, no. 5, pp. 28-34, Oct 2000.
[2] Trevor F. Cox and Michael A.A. Cox, "Metric Multidimensional Scaling," in Multidimensional Scaling. USA: Chapman & Hall/CRC, 2001, ch. 2, pp. 32-44.
[3] A. Boukerche, H.A.B. Oliveira, E.F. Nakamura, and A.A.F. Loureiro, "Localization systems for wireless sensor networks," Wireless Communications, IEEE , vol. 14, no. 6, pp. 6-12, December 2007.
[4] Wikipedia. (2011, July) Radar - Wikipedia, the free encyclopedia. [Online]. http://en.wikipedia.org/wiki/Radar
[5] K. Whitehouse, C. Karlof, and D. Ceuller, "A practical evaluation of radio signal strength for ranging-based localization," ACM SIGMOBILE Mobile Computing and Communications Review, vol. 11, no. 1, January 2007.
[6] A. Awad, T. Frunzke, and F. Dressler, "Adaptive Distance Estimation and Localization," in Digital System Design Architectures, Methods and Tools, 2007. DSD 2007. 10th Euromicro Conference, Lubeck, 2007, pp. 471-478.
[7] HowStuffWorks. (2011, July) How Wireless Mesh Networks Work. [Online]. http://communication.howstuffworks.com/how-wireless-mesh-networks-work.htm
[8] ZigBee Alliance. (2011, June) ZigBee Specifications. [Online]. http://www.zigbee.org/Specifications.aspx
[9] P. Besl and N. McKay, "A method for Registration of 3-D shapes," IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), vol. 14, no. 2, pp. 239 - 256, February 1992.
[10] K. Aamodt. (2010, Dec.) Texas Instruments. [Online]. http://focus.ti.com/lit/an/swra095/swra095.pdf
[11] J. Xu, M. Ma, and C.L. Law, "Position Estimation Using UWB TDOA Measurements," in Ultra-Wideband, The 2006 IEEE 2006 International Conference on, Waltham, MA, 2006, pp. 605-610.
[12] (2011, June) antenna-theory.com. [Online]. http://www.antenna-theory.com/basics/friis.php
[13] Guoqiang Mao, Barış Fidan, and Brian D. O. Anderson, "Wireless sensor network localization techniques, Computer Networks," The International Journal of Computer and Telecommunications Networking, vol. 51, no. 10, pp. 2529-2553, July 2007.
[14] M. Bal, M. Liu, W. Shen, and H. Ghenniwa, "Localization In Cooperative Wireless Sensor Network: A Review," in International Conference on Computer Supported Cooperative Work in Design, Santiago, Chile, 2009.
[15] R. Mautz, "Overview of Current Indoor Positioning Systems," in Geodezija
120
Ir Kartografija / Geodesy and Cartography, 2009 35(1), pp. 18-22.
[16] D. Niculescu and Badri Nath, "Ad hoc positioning system (APS) using AOA," in INFOCOM 2003. Twenty-Second Annual Joint Conference of the IEEE Computer and Communications. IEEE Societies, 30 March-3 April 2003 , pp. 1734 - 1743 vol.3.
[17] K.W. Cheung, H.C. So, W.K. Ma, and Y.T. Chan, "Received signal strength based mobile positioning via constrained weighted least squares," in Acoustics, Speech, and Signal Processing, 2003. Proceedings. (ICASSP '03). 2003 IEEE International Conference, 2003, pp. 137-140.
[18] Y. Kwon, K. Mechitov, S. Sundresh, W. Kim, and G. Agha, "Resilient localization for sensor networks in outdoor environments," Department of Computer Science, University of Illinois at Urbana Champaign, Technical Report UIUCDCS-R-2004-2449 March 2004.
[19] C.H. Wu, W. Sheng, and Y. Zhang, "Mobile Self-Localization using Multi-Dimensional Scaling in Robotic Sensor Networks," in IEEE International Conference on Robotics and Automation, 2007.
[20] Y. Shang and W. Ruml, "Improved MDS-Based Localization," in Proc. IEEE Infocom, 2004.
[21] Gopakumar Aloor and Lillykutty Jacob, "Distributed wireless sensor network localization using stochastic proximity embedding," Computer Communications, vol. 33, no. 6, pp. 745-755, April 2010.
[22] S. Capkun and J.P. Hubaux, "Secure positioning of wireless devices with application to sensor networks," in In IEEE INFOCOM, 2005.
[23] Lingfei Wu et al., "A practical evaluation of radio signal strength for mobile robot localization," in Robotics and Biomimetics (ROBIO), 2009 IEEE International Conference, Guilin, 19-23 Dec. 2009, pp. 516-522.
[24] P. Bahl and V.N. Padmanabhan, "RADAR: an in-building RF-based user location and tracking system," INFOCOM 2000. Nineteenth Annual Joint Conference of the IEEE Computer and Communications Societies. Proceedings. IEEE , vol. 2, pp. 775-784, August 2002.
[25] M. Bertinato et al., "RF localization and tracking of mobile nodes in wireless sensors networks:Architectures, algorithms and experiments," University of Padova, Technical report 2008.
[26] Q. Chen, D.L. Lee, and W.C. Lee, "Rule-based WiFi localization methods," in IEEE/IFIP international conference on embedded and ubiquitous computing, 2008.
[27] S. Ray, W. Lai, and I.Ch. Paschalidis, "Deployment optimization of sensornet-based stochastic location-detection systems," INFOCOM 2005. 24th Annual Joint Conference of the IEEE Computer and Communications Societies. Proceedings IEEE , vol. 4, pp. 13-17, August 2005.
[28] Fuxiang Gong, Wang Qing, and Xiaoguo Zhang, "A new distance based algorithm for TDOA localization in cellular networks," in Computer Science and Information Technology (ICCSIT), 2010 3rd IEEE International
121
Conference, Chengdu, 2010, pp. 502-505.
[29] Ran Yu, Bin Xu, Guodong Sun, and Zheng Yang, "Whistle: synchronization-free TDOA for localization," in Proceedings of the 8th ACM Conference on Embedded Networked Sensor Systems, 2010.
[30] Xing Wei, Ling Wang, and Jianwei Wan, "A New Localization Technique Based on Network TDOA Information," in ITS Telecommunications Proceedings, 2006 6th International Conference, Chengdu, 2007, pp. 127-130.
[31] Nissanka Bodhi Priyantha and Hari Balakrishnan, "The cricket indoor location system," Massachusetts Institute of Technology, Cambridge, MA, 2005.
[32] N.B. Priyantha, H. Balakrishnan, E. Demaine, and S. Teller, "Anchor-Free Distributed Localization in Sensor Networks," MIT Laboratory for Computer Science, Boston, Technical Report 2003.
[33] Rong Peng and Mihail L. Sichitiu, "Angle of Arrival Localization for Wireless Sensor Networks," in Sensor and Ad Hoc Communications and Networks, 2006. SECON '06. 2006 3rd Annual IEEE Communications Society, Reston, VA, 2007, pp. 374 - 382.
[34] ChulYoung Park et al., "AoA Localization System Design and Implementation Based on Zigbee for Applying Greenhouse," in Embedded and Multimedia Computing (EMC), 2010 5th International Conference, Cebu, 2010, pp. 1 - 4.
[35] Jose A. Costa, Neal Patwari, and III, Alfred O. Hero, "Distributed weighted-multidimensional scaling for node localization," ACM Transactions on Sensor Networks (TOSN), vol. 2, no. 1, February 2006.
[36] Yi Shang, Wheeler Ruml, Ying Zhang, and Markus Fromherz, "Localization from Connectivity in Sensor Networks," IEEE Transactions on Parallel and Distributed Systems, vol. 15, no. 11, pp. 961-974, November 2004.
[37] Jaehun Lee, WooYong Chung, and Euntai Kim, "Efficient range-free localization algorithm for wireless sensor network based on shortest path information," in ICCAS-SICE, 2009, Fukuoka, 2009, pp. 1962-1965.
[38] Dragoş Niculescu and Badri Nath, "DV Based Positioning in Ad Hoc Networks," in Telecommunication Systems, 2003, pp. 267-280.
[39] Hongyang Chen, Kaoru Sezaki, Ping Deng, and Hing Cheung So, "An improved DV-Hop localization algorithm for wireless sensor networks," in Industrial Electronics and Applications, 2008. ICIEA 2008. 3rd IEEE Conference, 3-5 June, 2008, pp. 1557-1561.
[40] T. He, C. Huang, B.M. Blum, J.A. Stankovic, and T.F. Abdelzaher, "Range-free localization schemes for large scale sensor networks," in International Conference on Mobile Computing and Networking, San Diego, CA, USA, 2003, pp. 81-95.
[41] Hyuk Lim and J.C. Hou, "Localization for anisotropic sensor networks," in INFOCOM 2005. 24th Annual Joint Conference of the IEEE Computer and
122
Communications Societies. Proceedings IEEE, 2005, pp. 138-149.
[42] S. Tian, X.M. Zhang, P.X. Liu, P. Sun, and X.G. Wang, "A RSSI-based DV-hop algorithm for wireless sensor networks," in Proc. of IEEE WICOM2007, Sep 2007.
[43] Bang Wang, Hock Beng Lim, Di Ma, and Cheng Fu, "The hop count shift problem and its impacts on protocol design in wireless ad hoc networks," Telecommunication Systems, vol. 44, pp. 49-60, November 2009.
[44] Benny S. L. Hung, "Wireless Mesh Network System Design and Implementation For Surveillance, Health Monitoring and Tracking," Simon Fraser University, Burnaby, Masters Thesis 2010.
[45] Antenova. (2011, June) Antenova - Integrated Antenna Solutions. [Online]. http://pdf.eicom.ru/datasheets/antenova/titanis_24_2010byyyy/titanis_24_2010byyyy.pdf
[46] D. Moore, J. Leonard, D. Rus, and S. Teller, "Robust distributed network localization with noisy range measurements," in Proc. SenSys., 2004, pp. 50-61.
[47] S. Basagni, "Distributed Clustering Algorithm for Ad-Hoc Networks," in Proc. Int'l Symp. Parallel Architectures, Algorithms, and Networks (I-SPAN), 1999.
[48] R. Krishnan and D. Starobinski, "Efficient clustering algorithms for self-organizing wireless sensor networks," J. Ad Hoc Networks, vol. 4, p. 36, 2006.
[49] Y. Zhou and L. Lamont, "An Optimal Local Map Registration Technique for Wireless Sensor Network Localization Problems," in Fusion, 2008.
[50] H. Chan, M. Luk, and A. Perrig, "Using clustering information for sensor network localization," in Proc. IEEE Conf. Distrib. Comput. Sensor Syst., 2005.
[51] O.H. Kwon and H.J. Song, "Localization through map stitching in wireless sensor networks," IEEE Trans. Parallel Distrib. Syst., vol. 19, no. 1, pp. 93-105, January 2008.
[52] R. Nagpal, "Organizing a global coordinate system from local information on an amorphous computer," MIT Department of Electrical Engineering and Computer Science, Boston, Technical Report 1999.
[53] H. Wymeersch, J. Lien, and M.Z. Win, "Cooperative localization in wireless networks," in Proc. IEEE (Special Issue on UWB Technology and Emerging Applications), 2009.
[54] Wikipedia. (2011, July) World Geodetic System - Wikipedia, the free encyclopedia. [Online]. http://en.wikipedia.org/wiki/World_Geodetic_System
